Pushing 1
Browse files- .gitattributes +1 -0
- .ipynb_checkpoints/__init__-checkpoint.py +0 -0
- .ipynb_checkpoints/modeling_feynmodel-checkpoint.py +1528 -0
- __init__.py +0 -0
- config.json +253 -0
- configuration_feynmodel.py +159 -0
- generation_config.json +11 -0
- model-00001-of-00003.safetensors +3 -0
- model-00002-of-00003.safetensors +3 -0
- model-00003-of-00003.safetensors +3 -0
- model.safetensors.index.json +705 -0
- modeling_feynmodel.py +1528 -0
- preprocessor_config.json +33 -0
- processing_florence2.py +1088 -0
- processor_config.json +6 -0
- special_tokens_map.json +30 -0
- tokenizer.json +3 -0
- tokenizer_config.json +2010 -0
.gitattributes
CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
.ipynb_checkpoints/__init__-checkpoint.py
ADDED
File without changes
|
.ipynb_checkpoints/modeling_feynmodel-checkpoint.py
ADDED
@@ -0,0 +1,1528 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# modeling_fynmodel : Imed MAGROUNE / 2024 - 09
|
2 |
+
# original code from modeling_FeynModel
|
3 |
+
# add DaVit Vision Tower
|
4 |
+
#
|
5 |
+
# update generate forward function
|
6 |
+
#
|
7 |
+
# add lora adapters
|
8 |
+
#
|
9 |
+
# train on coco OD and vision reasoning
|
10 |
+
# train on ScenceQA
|
11 |
+
#
|
12 |
+
# todo add mamaba layer
|
13 |
+
#
|
14 |
+
# todo train on Arc-AGI
|
15 |
+
|
16 |
+
|
17 |
+
from transformers.modeling_utils import PreTrainedModel
|
18 |
+
from transformers.utils import (
|
19 |
+
ModelOutput,
|
20 |
+
add_start_docstrings,
|
21 |
+
add_start_docstrings_to_model_forward,
|
22 |
+
is_flash_attn_2_available,
|
23 |
+
logging,
|
24 |
+
replace_return_docstrings,
|
25 |
+
is_flash_attn_2_available,
|
26 |
+
is_flash_attn_greater_or_equal_2_10,
|
27 |
+
)
|
28 |
+
from transformers.activations import ACT2FN
|
29 |
+
from transformers.modeling_attn_mask_utils import (
|
30 |
+
_prepare_4d_attention_mask,
|
31 |
+
_prepare_4d_attention_mask_for_sdpa,
|
32 |
+
_prepare_4d_causal_attention_mask,
|
33 |
+
_prepare_4d_causal_attention_mask_for_sdpa,
|
34 |
+
)
|
35 |
+
from transformers.modeling_outputs import (
|
36 |
+
BaseModelOutput,
|
37 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
38 |
+
Seq2SeqLMOutput,
|
39 |
+
Seq2SeqModelOutput,
|
40 |
+
)
|
41 |
+
|
42 |
+
from transformers.cache_utils import Cache, HybridCache
|
43 |
+
from transformers.modeling_outputs import (
|
44 |
+
BaseModelOutputWithPast,
|
45 |
+
CausalLMOutputWithPast,
|
46 |
+
SequenceClassifierOutputWithPast,
|
47 |
+
TokenClassifierOutput,
|
48 |
+
)
|
49 |
+
|
50 |
+
from typing import List, Optional, Tuple, Union
|
51 |
+
|
52 |
+
from transformers.models.gemma2.modeling_gemma2 import Gemma2Model, Gemma2ForCausalLM,Gemma2DecoderLayer,Gemma2RMSNorm
|
53 |
+
from configuration_feynmodel import FeynModelConfig,Florence2VisionConfig
|
54 |
+
|
55 |
+
from transformers import AutoProcessor, AutoTokenizer, AutoModelForCausalLM
|
56 |
+
import json
|
57 |
+
import math
|
58 |
+
import torch
|
59 |
+
from torch import nn
|
60 |
+
import torch.nn.functional as F
|
61 |
+
import logging
|
62 |
+
|
63 |
+
from transformers.utils import (
|
64 |
+
ModelOutput,
|
65 |
+
add_start_docstrings,
|
66 |
+
add_start_docstrings_to_model_forward,
|
67 |
+
is_flash_attn_2_available,
|
68 |
+
logging,
|
69 |
+
replace_return_docstrings,
|
70 |
+
is_flash_attn_2_available,
|
71 |
+
is_flash_attn_greater_or_equal_2_10,
|
72 |
+
)
|
73 |
+
|
74 |
+
from transformers.modeling_utils import PreTrainedModel
|
75 |
+
|
76 |
+
from collections import OrderedDict
|
77 |
+
from einops import rearrange
|
78 |
+
from timm.models.layers import DropPath, trunc_normal_
|
79 |
+
|
80 |
+
logger = logging.get_logger(__name__)
|
81 |
+
|
82 |
+
class MySequential(nn.Sequential):
|
83 |
+
def forward(self, *inputs):
|
84 |
+
for module in self._modules.values():
|
85 |
+
if type(inputs) == tuple:
|
86 |
+
inputs = module(*inputs)
|
87 |
+
else:
|
88 |
+
inputs = module(inputs)
|
89 |
+
return inputs
|
90 |
+
|
91 |
+
|
92 |
+
class PreNorm(nn.Module):
|
93 |
+
def __init__(self, norm, fn, drop_path=None):
|
94 |
+
super().__init__()
|
95 |
+
self.norm = norm
|
96 |
+
self.fn = fn
|
97 |
+
self.drop_path = drop_path
|
98 |
+
|
99 |
+
def forward(self, x, *args, **kwargs):
|
100 |
+
shortcut = x
|
101 |
+
if self.norm != None:
|
102 |
+
x, size = self.fn(self.norm(x), *args, **kwargs)
|
103 |
+
else:
|
104 |
+
x, size = self.fn(x, *args, **kwargs)
|
105 |
+
|
106 |
+
if self.drop_path:
|
107 |
+
x = self.drop_path(x)
|
108 |
+
|
109 |
+
x = shortcut + x
|
110 |
+
|
111 |
+
return x, size
|
112 |
+
|
113 |
+
|
114 |
+
class Mlp(nn.Module):
|
115 |
+
def __init__(
|
116 |
+
self,
|
117 |
+
in_features,
|
118 |
+
hidden_features=None,
|
119 |
+
out_features=None,
|
120 |
+
act_layer=nn.GELU,
|
121 |
+
):
|
122 |
+
super().__init__()
|
123 |
+
out_features = out_features or in_features
|
124 |
+
hidden_features = hidden_features or in_features
|
125 |
+
self.net = nn.Sequential(OrderedDict([
|
126 |
+
("fc1", nn.Linear(in_features, hidden_features)),
|
127 |
+
("act", act_layer()),
|
128 |
+
("fc2", nn.Linear(hidden_features, out_features))
|
129 |
+
]))
|
130 |
+
|
131 |
+
def forward(self, x, size):
|
132 |
+
return self.net(x), size
|
133 |
+
|
134 |
+
|
135 |
+
class DepthWiseConv2d(nn.Module):
|
136 |
+
def __init__(
|
137 |
+
self,
|
138 |
+
dim_in,
|
139 |
+
kernel_size,
|
140 |
+
padding,
|
141 |
+
stride,
|
142 |
+
bias=True,
|
143 |
+
):
|
144 |
+
super().__init__()
|
145 |
+
self.dw = nn.Conv2d(
|
146 |
+
dim_in, dim_in,
|
147 |
+
kernel_size=kernel_size,
|
148 |
+
padding=padding,
|
149 |
+
groups=dim_in,
|
150 |
+
stride=stride,
|
151 |
+
bias=bias
|
152 |
+
)
|
153 |
+
|
154 |
+
def forward(self, x, size):
|
155 |
+
B, N, C = x.shape
|
156 |
+
H, W = size
|
157 |
+
assert N == H * W
|
158 |
+
|
159 |
+
x = self.dw(x.transpose(1, 2).view(B, C, H, W))
|
160 |
+
size = (x.size(-2), x.size(-1))
|
161 |
+
x = x.flatten(2).transpose(1, 2)
|
162 |
+
return x, size
|
163 |
+
|
164 |
+
|
165 |
+
class ConvEmbed(nn.Module):
|
166 |
+
""" Image to Patch Embedding
|
167 |
+
"""
|
168 |
+
|
169 |
+
def __init__(
|
170 |
+
self,
|
171 |
+
patch_size=7,
|
172 |
+
in_chans=3,
|
173 |
+
embed_dim=64,
|
174 |
+
stride=4,
|
175 |
+
padding=2,
|
176 |
+
norm_layer=None,
|
177 |
+
pre_norm=True
|
178 |
+
):
|
179 |
+
super().__init__()
|
180 |
+
self.patch_size = patch_size
|
181 |
+
|
182 |
+
self.proj = nn.Conv2d(
|
183 |
+
in_chans, embed_dim,
|
184 |
+
kernel_size=patch_size,
|
185 |
+
stride=stride,
|
186 |
+
padding=padding
|
187 |
+
)
|
188 |
+
|
189 |
+
dim_norm = in_chans if pre_norm else embed_dim
|
190 |
+
self.norm = norm_layer(dim_norm) if norm_layer else None
|
191 |
+
|
192 |
+
self.pre_norm = pre_norm
|
193 |
+
|
194 |
+
def forward(self, x, size):
|
195 |
+
H, W = size
|
196 |
+
if len(x.size()) == 3:
|
197 |
+
if self.norm and self.pre_norm:
|
198 |
+
x = self.norm(x)
|
199 |
+
x = rearrange(
|
200 |
+
x, 'b (h w) c -> b c h w',
|
201 |
+
h=H, w=W
|
202 |
+
)
|
203 |
+
|
204 |
+
x = self.proj(x)
|
205 |
+
|
206 |
+
_, _, H, W = x.shape
|
207 |
+
x = rearrange(x, 'b c h w -> b (h w) c')
|
208 |
+
if self.norm and not self.pre_norm:
|
209 |
+
x = self.norm(x)
|
210 |
+
|
211 |
+
return x, (H, W)
|
212 |
+
|
213 |
+
|
214 |
+
class ChannelAttention(nn.Module):
|
215 |
+
|
216 |
+
def __init__(self, dim, groups=8, qkv_bias=True):
|
217 |
+
super().__init__()
|
218 |
+
|
219 |
+
self.groups = groups
|
220 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
221 |
+
self.proj = nn.Linear(dim, dim)
|
222 |
+
|
223 |
+
def forward(self, x, size):
|
224 |
+
B, N, C = x.shape
|
225 |
+
|
226 |
+
qkv = self.qkv(x).reshape(B, N, 3, self.groups, C // self.groups).permute(2, 0, 3, 1, 4)
|
227 |
+
q, k, v = qkv[0], qkv[1], qkv[2]
|
228 |
+
|
229 |
+
q = q * (float(N) ** -0.5)
|
230 |
+
attention = q.transpose(-1, -2) @ k
|
231 |
+
attention = attention.softmax(dim=-1)
|
232 |
+
x = (attention @ v.transpose(-1, -2)).transpose(-1, -2)
|
233 |
+
x = x.transpose(1, 2).reshape(B, N, C)
|
234 |
+
x = self.proj(x)
|
235 |
+
return x, size
|
236 |
+
|
237 |
+
|
238 |
+
class ChannelBlock(nn.Module):
|
239 |
+
|
240 |
+
def __init__(self, dim, groups, mlp_ratio=4., qkv_bias=True,
|
241 |
+
drop_path_rate=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm,
|
242 |
+
conv_at_attn=True, conv_at_ffn=True):
|
243 |
+
super().__init__()
|
244 |
+
|
245 |
+
drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
|
246 |
+
|
247 |
+
self.conv1 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_attn else None
|
248 |
+
self.channel_attn = PreNorm(
|
249 |
+
norm_layer(dim),
|
250 |
+
ChannelAttention(dim, groups=groups, qkv_bias=qkv_bias),
|
251 |
+
drop_path
|
252 |
+
)
|
253 |
+
self.conv2 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_ffn else None
|
254 |
+
self.ffn = PreNorm(
|
255 |
+
norm_layer(dim),
|
256 |
+
Mlp(in_features=dim, hidden_features=int(dim*mlp_ratio), act_layer=act_layer),
|
257 |
+
drop_path
|
258 |
+
)
|
259 |
+
|
260 |
+
def forward(self, x, size):
|
261 |
+
if self.conv1:
|
262 |
+
x, size = self.conv1(x, size)
|
263 |
+
x, size = self.channel_attn(x, size)
|
264 |
+
|
265 |
+
if self.conv2:
|
266 |
+
x, size = self.conv2(x, size)
|
267 |
+
x, size = self.ffn(x, size)
|
268 |
+
|
269 |
+
return x, size
|
270 |
+
|
271 |
+
|
272 |
+
def window_partition(x, window_size: int):
|
273 |
+
B, H, W, C = x.shape
|
274 |
+
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
|
275 |
+
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
|
276 |
+
return windows
|
277 |
+
|
278 |
+
|
279 |
+
def window_reverse(windows, batch_size: int, window_size: int, H: int, W: int):
|
280 |
+
B = batch_size
|
281 |
+
# this will cause onnx conversion failed for dynamic axis, because treated as constant
|
282 |
+
# int(windows.shape[0] / (H * W / window_size / window_size))
|
283 |
+
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
|
284 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
|
285 |
+
return x
|
286 |
+
|
287 |
+
|
288 |
+
class WindowAttention(nn.Module):
|
289 |
+
def __init__(self, dim, num_heads, window_size, qkv_bias=True):
|
290 |
+
|
291 |
+
super().__init__()
|
292 |
+
self.dim = dim
|
293 |
+
self.window_size = window_size
|
294 |
+
self.num_heads = num_heads
|
295 |
+
head_dim = dim // num_heads
|
296 |
+
self.scale = float(head_dim) ** -0.5
|
297 |
+
|
298 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
299 |
+
self.proj = nn.Linear(dim, dim)
|
300 |
+
|
301 |
+
self.softmax = nn.Softmax(dim=-1)
|
302 |
+
|
303 |
+
def forward(self, x, size):
|
304 |
+
|
305 |
+
H, W = size
|
306 |
+
B, L, C = x.shape
|
307 |
+
assert L == H * W, "input feature has wrong size"
|
308 |
+
|
309 |
+
x = x.view(B, H, W, C)
|
310 |
+
|
311 |
+
pad_l = pad_t = 0
|
312 |
+
pad_r = (self.window_size - W % self.window_size) % self.window_size
|
313 |
+
pad_b = (self.window_size - H % self.window_size) % self.window_size
|
314 |
+
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
|
315 |
+
_, Hp, Wp, _ = x.shape
|
316 |
+
|
317 |
+
x = window_partition(x, self.window_size)
|
318 |
+
x = x.view(-1, self.window_size * self.window_size, C)
|
319 |
+
|
320 |
+
# W-MSA/SW-MSA
|
321 |
+
# attn_windows = self.attn(x_windows)
|
322 |
+
|
323 |
+
B_, N, C = x.shape
|
324 |
+
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
325 |
+
q, k, v = qkv[0], qkv[1], qkv[2]
|
326 |
+
|
327 |
+
q = q * self.scale
|
328 |
+
attn = (q @ k.transpose(-2, -1))
|
329 |
+
attn = self.softmax(attn)
|
330 |
+
|
331 |
+
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
|
332 |
+
x = self.proj(x)
|
333 |
+
|
334 |
+
# merge windows
|
335 |
+
x = x.view(
|
336 |
+
-1, self.window_size, self.window_size, C
|
337 |
+
)
|
338 |
+
x = window_reverse(x, B, self.window_size, Hp, Wp)
|
339 |
+
|
340 |
+
if pad_r > 0 or pad_b > 0:
|
341 |
+
x = x[:, :H, :W, :].contiguous()
|
342 |
+
|
343 |
+
x = x.view(B, H * W, C)
|
344 |
+
|
345 |
+
return x, size
|
346 |
+
|
347 |
+
|
348 |
+
class SpatialBlock(nn.Module):
|
349 |
+
|
350 |
+
def __init__(self, dim, num_heads, window_size,
|
351 |
+
mlp_ratio=4., qkv_bias=True, drop_path_rate=0., act_layer=nn.GELU,
|
352 |
+
norm_layer=nn.LayerNorm, conv_at_attn=True, conv_at_ffn=True):
|
353 |
+
super().__init__()
|
354 |
+
|
355 |
+
drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
|
356 |
+
|
357 |
+
self.conv1 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_attn else None
|
358 |
+
self.window_attn = PreNorm(
|
359 |
+
norm_layer(dim),
|
360 |
+
WindowAttention(dim, num_heads, window_size, qkv_bias=qkv_bias),
|
361 |
+
drop_path
|
362 |
+
)
|
363 |
+
self.conv2 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_ffn else None
|
364 |
+
self.ffn = PreNorm(
|
365 |
+
norm_layer(dim),
|
366 |
+
Mlp(in_features=dim, hidden_features=int(dim*mlp_ratio), act_layer=act_layer),
|
367 |
+
drop_path
|
368 |
+
)
|
369 |
+
|
370 |
+
def forward(self, x, size):
|
371 |
+
if self.conv1:
|
372 |
+
x, size = self.conv1(x, size)
|
373 |
+
x, size = self.window_attn(x, size)
|
374 |
+
|
375 |
+
if self.conv2:
|
376 |
+
x, size = self.conv2(x, size)
|
377 |
+
x, size = self.ffn(x, size)
|
378 |
+
return x, size
|
379 |
+
|
380 |
+
|
381 |
+
class DaViT(nn.Module):
|
382 |
+
""" DaViT: Dual-Attention Transformer
|
383 |
+
|
384 |
+
Args:
|
385 |
+
in_chans (int): Number of input image channels. Default: 3.
|
386 |
+
num_classes (int): Number of classes for classification head. Default: 1000.
|
387 |
+
patch_size (tuple(int)): Patch size of convolution in different stages. Default: (7, 2, 2, 2).
|
388 |
+
patch_stride (tuple(int)): Patch stride of convolution in different stages. Default: (4, 2, 2, 2).
|
389 |
+
patch_padding (tuple(int)): Patch padding of convolution in different stages. Default: (3, 0, 0, 0).
|
390 |
+
patch_prenorm (tuple(bool)): If True, perform norm before convlution layer. Default: (True, False, False, False).
|
391 |
+
embed_dims (tuple(int)): Patch embedding dimension in different stages. Default: (64, 128, 192, 256).
|
392 |
+
num_heads (tuple(int)): Number of spatial attention heads in different stages. Default: (4, 8, 12, 16).
|
393 |
+
num_groups (tuple(int)): Number of channel groups in different stages. Default: (4, 8, 12, 16).
|
394 |
+
window_size (int): Window size. Default: 7.
|
395 |
+
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
|
396 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True.
|
397 |
+
drop_path_rate (float): Stochastic depth rate. Default: 0.1.
|
398 |
+
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
|
399 |
+
enable_checkpoint (bool): If True, enable checkpointing. Default: False.
|
400 |
+
conv_at_attn (bool): If True, performe depthwise convolution before attention layer. Default: True.
|
401 |
+
conv_at_ffn (bool): If True, performe depthwise convolution before ffn layer. Default: True.
|
402 |
+
"""
|
403 |
+
|
404 |
+
def __init__(
|
405 |
+
self,
|
406 |
+
in_chans=3,
|
407 |
+
num_classes=1000,
|
408 |
+
depths=(1, 1, 3, 1),
|
409 |
+
patch_size=(7, 2, 2, 2),
|
410 |
+
patch_stride=(4, 2, 2, 2),
|
411 |
+
patch_padding=(3, 0, 0, 0),
|
412 |
+
patch_prenorm=(False, False, False, False),
|
413 |
+
embed_dims=(64, 128, 192, 256),
|
414 |
+
num_heads=(3, 6, 12, 24),
|
415 |
+
num_groups=(3, 6, 12, 24),
|
416 |
+
window_size=7,
|
417 |
+
mlp_ratio=4.,
|
418 |
+
qkv_bias=True,
|
419 |
+
drop_path_rate=0.1,
|
420 |
+
norm_layer=nn.LayerNorm,
|
421 |
+
enable_checkpoint=False,
|
422 |
+
conv_at_attn=True,
|
423 |
+
conv_at_ffn=True,
|
424 |
+
):
|
425 |
+
super().__init__()
|
426 |
+
|
427 |
+
self.num_classes = num_classes
|
428 |
+
self.embed_dims = embed_dims
|
429 |
+
self.num_heads = num_heads
|
430 |
+
self.num_groups = num_groups
|
431 |
+
self.num_stages = len(self.embed_dims)
|
432 |
+
self.enable_checkpoint = enable_checkpoint
|
433 |
+
assert self.num_stages == len(self.num_heads) == len(self.num_groups)
|
434 |
+
|
435 |
+
num_stages = len(embed_dims)
|
436 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths)*2)]
|
437 |
+
|
438 |
+
depth_offset = 0
|
439 |
+
convs = []
|
440 |
+
blocks = []
|
441 |
+
for i in range(num_stages):
|
442 |
+
conv_embed = ConvEmbed(
|
443 |
+
patch_size=patch_size[i],
|
444 |
+
stride=patch_stride[i],
|
445 |
+
padding=patch_padding[i],
|
446 |
+
in_chans=in_chans if i == 0 else self.embed_dims[i - 1],
|
447 |
+
embed_dim=self.embed_dims[i],
|
448 |
+
norm_layer=norm_layer,
|
449 |
+
pre_norm=patch_prenorm[i]
|
450 |
+
)
|
451 |
+
convs.append(conv_embed)
|
452 |
+
|
453 |
+
block = MySequential(
|
454 |
+
*[
|
455 |
+
MySequential(OrderedDict([
|
456 |
+
(
|
457 |
+
'spatial_block', SpatialBlock(
|
458 |
+
embed_dims[i],
|
459 |
+
num_heads[i],
|
460 |
+
window_size,
|
461 |
+
drop_path_rate=dpr[depth_offset+j*2],
|
462 |
+
qkv_bias=qkv_bias,
|
463 |
+
mlp_ratio=mlp_ratio,
|
464 |
+
conv_at_attn=conv_at_attn,
|
465 |
+
conv_at_ffn=conv_at_ffn,
|
466 |
+
)
|
467 |
+
),
|
468 |
+
(
|
469 |
+
'channel_block', ChannelBlock(
|
470 |
+
embed_dims[i],
|
471 |
+
num_groups[i],
|
472 |
+
drop_path_rate=dpr[depth_offset+j*2+1],
|
473 |
+
qkv_bias=qkv_bias,
|
474 |
+
mlp_ratio=mlp_ratio,
|
475 |
+
conv_at_attn=conv_at_attn,
|
476 |
+
conv_at_ffn=conv_at_ffn,
|
477 |
+
)
|
478 |
+
)
|
479 |
+
])) for j in range(depths[i])
|
480 |
+
]
|
481 |
+
)
|
482 |
+
blocks.append(block)
|
483 |
+
depth_offset += depths[i]*2
|
484 |
+
|
485 |
+
self.convs = nn.ModuleList(convs)
|
486 |
+
self.blocks = nn.ModuleList(blocks)
|
487 |
+
|
488 |
+
self.norms = norm_layer(self.embed_dims[-1])
|
489 |
+
self.avgpool = nn.AdaptiveAvgPool1d(1)
|
490 |
+
self.head = nn.Linear(self.embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity()
|
491 |
+
|
492 |
+
self.apply(self._init_weights)
|
493 |
+
|
494 |
+
@property
|
495 |
+
def dim_out(self):
|
496 |
+
return self.embed_dims[-1]
|
497 |
+
|
498 |
+
def _init_weights(self, m):
|
499 |
+
if isinstance(m, nn.Linear):
|
500 |
+
trunc_normal_(m.weight, std=0.02)
|
501 |
+
if m.bias is not None:
|
502 |
+
nn.init.constant_(m.bias, 0)
|
503 |
+
elif isinstance(m, nn.Conv2d):
|
504 |
+
nn.init.normal_(m.weight, std=0.02)
|
505 |
+
for name, _ in m.named_parameters():
|
506 |
+
if name in ['bias']:
|
507 |
+
nn.init.constant_(m.bias, 0)
|
508 |
+
elif isinstance(m, nn.LayerNorm):
|
509 |
+
nn.init.constant_(m.weight, 1.0)
|
510 |
+
nn.init.constant_(m.bias, 0)
|
511 |
+
elif isinstance(m, nn.BatchNorm2d):
|
512 |
+
nn.init.constant_(m.weight, 1.0)
|
513 |
+
nn.init.constant_(m.bias, 0)
|
514 |
+
|
515 |
+
def forward_features_unpool(self, x):
|
516 |
+
"""
|
517 |
+
forward until avg pooling
|
518 |
+
Args:
|
519 |
+
x (_type_): input image tensor
|
520 |
+
"""
|
521 |
+
input_size = (x.size(2), x.size(3))
|
522 |
+
for conv, block in zip(self.convs, self.blocks):
|
523 |
+
x, input_size = conv(x, input_size)
|
524 |
+
if self.enable_checkpoint:
|
525 |
+
x, input_size = checkpoint.checkpoint(block, x, input_size)
|
526 |
+
else:
|
527 |
+
x, input_size = block(x, input_size)
|
528 |
+
return x
|
529 |
+
|
530 |
+
def forward_features(self, x):
|
531 |
+
x = self.forward_features_unpool(x)
|
532 |
+
|
533 |
+
# (batch_size, num_tokens, token_dim)
|
534 |
+
x = self.avgpool(x.transpose(1, 2))
|
535 |
+
# (batch_size, 1, num_tokens)
|
536 |
+
x = torch.flatten(x, 1)
|
537 |
+
x = self.norms(x)
|
538 |
+
|
539 |
+
return x
|
540 |
+
|
541 |
+
def forward(self, x):
|
542 |
+
x = self.forward_features(x)
|
543 |
+
x = self.head(x)
|
544 |
+
return x
|
545 |
+
|
546 |
+
@classmethod
|
547 |
+
def from_config(cls, config):
|
548 |
+
return cls(
|
549 |
+
depths=config.depths,
|
550 |
+
embed_dims=config.dim_embed,
|
551 |
+
num_heads=config.num_heads,
|
552 |
+
num_groups=config.num_groups,
|
553 |
+
patch_size=config.patch_size,
|
554 |
+
patch_stride=config.patch_stride,
|
555 |
+
patch_padding=config.patch_padding,
|
556 |
+
patch_prenorm=config.patch_prenorm,
|
557 |
+
drop_path_rate=config.drop_path_rate,
|
558 |
+
window_size=config.window_size,
|
559 |
+
)
|
560 |
+
|
561 |
+
|
562 |
+
|
563 |
+
|
564 |
+
_CONFIG_FOR_DOC = "FeynModelConfig"
|
565 |
+
|
566 |
+
FEYNMODEL_START_DOCSTRING = r"""
|
567 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
568 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
569 |
+
etc.)
|
570 |
+
|
571 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
572 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
573 |
+
and behavior.
|
574 |
+
|
575 |
+
Parameters:
|
576 |
+
config ([`FeynModelConfig`]):
|
577 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
578 |
+
load the weights associated with the model, only the configuration. Check out the
|
579 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
580 |
+
"""
|
581 |
+
FEYNMODEL_INPUTS_DOCSTRING = r"""
|
582 |
+
Args:
|
583 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
584 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
585 |
+
it.
|
586 |
+
|
587 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
588 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
589 |
+
|
590 |
+
[What are input IDs?](../glossary#input-ids)
|
591 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
592 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
593 |
+
|
594 |
+
- 1 for tokens that are **not masked**,
|
595 |
+
- 0 for tokens that are **masked**.
|
596 |
+
|
597 |
+
[What are attention masks?](../glossary#attention-mask)
|
598 |
+
|
599 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
600 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
601 |
+
|
602 |
+
If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
|
603 |
+
`past_key_values`).
|
604 |
+
|
605 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
606 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
607 |
+
information on the default strategy.
|
608 |
+
|
609 |
+
- 1 indicates the head is **not masked**,
|
610 |
+
- 0 indicates the head is **masked**.
|
611 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
612 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
613 |
+
config.n_positions - 1]`.
|
614 |
+
|
615 |
+
[What are position IDs?](../glossary#position-ids)
|
616 |
+
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
617 |
+
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
618 |
+
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
619 |
+
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
620 |
+
|
621 |
+
Two formats are allowed:
|
622 |
+
- a [`~cache_utils.Cache`] instance;
|
623 |
+
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
624 |
+
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
625 |
+
cache format.
|
626 |
+
|
627 |
+
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
628 |
+
legacy cache format will be returned.
|
629 |
+
|
630 |
+
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
631 |
+
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
632 |
+
of shape `(batch_size, sequence_length)`.
|
633 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
634 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
635 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
636 |
+
model's internal embedding lookup matrix.
|
637 |
+
use_cache (`bool`, *optional*):
|
638 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
639 |
+
`past_key_values`).
|
640 |
+
output_attentions (`bool`, *optional*):
|
641 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
642 |
+
tensors for more detail.
|
643 |
+
output_hidden_states (`bool`, *optional*):
|
644 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
645 |
+
more detail.
|
646 |
+
return_dict (`bool`, *optional*):
|
647 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
648 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
649 |
+
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
|
650 |
+
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
|
651 |
+
the complete sequence length.
|
652 |
+
"""
|
653 |
+
|
654 |
+
# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
|
655 |
+
def _prepare_4d_causal_attention_mask_with_cache_position(
|
656 |
+
attention_mask: torch.Tensor,
|
657 |
+
sequence_length: int,
|
658 |
+
target_length: int,
|
659 |
+
dtype: torch.dtype,
|
660 |
+
device: torch.device,
|
661 |
+
min_dtype: float,
|
662 |
+
cache_position: torch.Tensor,
|
663 |
+
batch_size: int,
|
664 |
+
):
|
665 |
+
|
666 |
+
#print(f" +++++++++ prepare 4K +++++++++++++++ rec {attention_mask.size()} sequence_length {sequence_length}")
|
667 |
+
if attention_mask is not None and attention_mask.dim() == 4:
|
668 |
+
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
669 |
+
#print("+++++++++++++++++ return it")
|
670 |
+
#causal_mask = attention_mask
|
671 |
+
# In this case we assume that the mask comes already in inverted form.
|
672 |
+
causal_mask = attention_mask[:, :, -sequence_length:, :]
|
673 |
+
#print(f"+++++++++++++++++ truncated causal_mask to last {sequence_length} elements, size: {causal_mask.size()}")
|
674 |
+
#print(f"+++++++++++++++++ return it causal_mask {causal_mask.size()} !!!!!!!!! attention_mask {attention_mask.size()}")
|
675 |
+
else:
|
676 |
+
#print("+++++++++++++++++++++ else +++++++++++++++++")
|
677 |
+
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
|
678 |
+
#print(f"++++++++++++++++ causal_mask {causal_mask.size()} ++++++++++++++++++ sequence_length = {sequence_length} ")
|
679 |
+
if sequence_length != 1:
|
680 |
+
causal_mask = torch.triu(causal_mask, diagonal=1)
|
681 |
+
#print(f"++++++++++++++++++ causal_mask = torch.triu ++++++++++ {causal_mask.size()} ")
|
682 |
+
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
683 |
+
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
684 |
+
#print(f"+++++++++++++++++++++ avant if attention_mask is not None:, causal_mask={causal_mask.size()}")
|
685 |
+
if attention_mask is not None:
|
686 |
+
#print(" +++++++++++++ attention_mask is None++++++++++++")
|
687 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
688 |
+
mask_length = attention_mask.shape[-1]
|
689 |
+
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
|
690 |
+
padding_mask = padding_mask == 0
|
691 |
+
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
692 |
+
padding_mask, min_dtype
|
693 |
+
)
|
694 |
+
#print(f"+++++++++++++++++++ 4K returning causal_mask {causal_mask.size()} +++++++++++++++++++")
|
695 |
+
|
696 |
+
return causal_mask
|
697 |
+
|
698 |
+
class LearnedAbsolutePositionEmbedding2D(nn.Module):
|
699 |
+
"""
|
700 |
+
This module learns positional embeddings up to a fixed maximum size.
|
701 |
+
"""
|
702 |
+
|
703 |
+
def __init__(self, embedding_dim=256, num_pos=50):
|
704 |
+
super().__init__()
|
705 |
+
self.row_embeddings = nn.Embedding(num_pos, embedding_dim // 2)
|
706 |
+
self.column_embeddings = nn.Embedding(num_pos, embedding_dim - (embedding_dim // 2))
|
707 |
+
|
708 |
+
def forward(self, pixel_values):
|
709 |
+
"""
|
710 |
+
pixel_values: (batch_size, height, width, num_channels)
|
711 |
+
returns: (batch_size, height, width, embedding_dim * 2)
|
712 |
+
"""
|
713 |
+
if len(pixel_values.shape) != 4:
|
714 |
+
raise ValueError('pixel_values must be a 4D tensor')
|
715 |
+
height, width = pixel_values.shape[1:3]
|
716 |
+
width_values = torch.arange(width, device=pixel_values.device)
|
717 |
+
height_values = torch.arange(height, device=pixel_values.device)
|
718 |
+
x_emb = self.column_embeddings(width_values)
|
719 |
+
y_emb = self.row_embeddings(height_values)
|
720 |
+
# (height, width, embedding_dim * 2)
|
721 |
+
pos = torch.cat([x_emb.unsqueeze(0).repeat(height, 1, 1), y_emb.unsqueeze(1).repeat(1, width, 1)], dim=-1)
|
722 |
+
# (embedding_dim * 2, height, width)
|
723 |
+
pos = pos.permute(2, 0, 1)
|
724 |
+
pos = pos.unsqueeze(0)
|
725 |
+
# (batch_size, embedding_dim * 2, height, width)
|
726 |
+
pos = pos.repeat(pixel_values.shape[0], 1, 1, 1)
|
727 |
+
# (batch_size, height, width, embedding_dim * 2)
|
728 |
+
pos = pos.permute(0, 2, 3, 1)
|
729 |
+
return pos
|
730 |
+
|
731 |
+
class PositionalEmbeddingCosine1D(nn.Module):
|
732 |
+
"""
|
733 |
+
This class implements a very simple positional encoding. It follows closely
|
734 |
+
the encoder from the link below:
|
735 |
+
https://pytorch.org/tutorials/beginner/translation_transformer.html
|
736 |
+
Args:
|
737 |
+
embed_dim: The dimension of the embeddings.
|
738 |
+
dropout_prob: The dropout probability.
|
739 |
+
max_seq_len: The maximum length to precompute the positional encodings.
|
740 |
+
"""
|
741 |
+
def __init__(
|
742 |
+
self,
|
743 |
+
embed_dim: int = 512,
|
744 |
+
max_seq_len: int = 1024) -> None:
|
745 |
+
super(PositionalEmbeddingCosine1D, self).__init__()
|
746 |
+
self.embed_dim = embed_dim
|
747 |
+
self.max_seq_len = max_seq_len
|
748 |
+
# Generate the sinusoidal arrays.
|
749 |
+
factor = math.log(10000)
|
750 |
+
denominator = torch.exp(
|
751 |
+
-factor * torch.arange(0, self.embed_dim, 2) / self.embed_dim)
|
752 |
+
# Matrix where rows correspond to a positional embedding as a function
|
753 |
+
# of the position index (i.e., the row index).
|
754 |
+
frequencies = \
|
755 |
+
torch.arange(0, self.max_seq_len) \
|
756 |
+
.reshape(self.max_seq_len, 1) * denominator
|
757 |
+
pos_idx_to_embed = torch.zeros((self.max_seq_len, self.embed_dim))
|
758 |
+
# Populate uneven entries.
|
759 |
+
pos_idx_to_embed[:, 0::2] = torch.sin(frequencies)
|
760 |
+
pos_idx_to_embed[:, 1::2] = torch.cos(frequencies)
|
761 |
+
# Save the positional embeddings in a constant buffer.
|
762 |
+
self.register_buffer("pos_idx_to_embed", pos_idx_to_embed)
|
763 |
+
|
764 |
+
def forward(self, seq_embeds: torch.Tensor) -> torch.Tensor:
|
765 |
+
"""
|
766 |
+
Args:
|
767 |
+
seq_embeds: The sequence embeddings in order. Allowed size:
|
768 |
+
1. [T, D], where T is the length of the sequence, and D is the
|
769 |
+
frame embedding dimension.
|
770 |
+
2. [B, T, D], where B is the batch size and T and D are the
|
771 |
+
same as above.
|
772 |
+
Returns a tensor of with the same dimensions as the input: i.e.,
|
773 |
+
[1, T, D] or [T, D].
|
774 |
+
"""
|
775 |
+
shape_len = len(seq_embeds.shape)
|
776 |
+
assert 2 <= shape_len <= 3
|
777 |
+
len_seq = seq_embeds.size(-2)
|
778 |
+
assert len_seq <= self.max_seq_len
|
779 |
+
pos_embeds = self.pos_idx_to_embed[0:seq_embeds.size(-2), :]
|
780 |
+
# Adapt pre-computed positional embeddings to the input.
|
781 |
+
if shape_len == 3:
|
782 |
+
pos_embeds = pos_embeds.view(
|
783 |
+
(1, pos_embeds.size(0), pos_embeds.size(1)))
|
784 |
+
return pos_embeds
|
785 |
+
|
786 |
+
|
787 |
+
class LearnedAbsolutePositionEmbedding1D(nn.Module):
|
788 |
+
"""
|
789 |
+
Learnable absolute positional embeddings for 1D sequences.
|
790 |
+
Args:
|
791 |
+
embed_dim: The dimension of the embeddings.
|
792 |
+
max_seq_len: The maximum length to precompute the positional encodings.
|
793 |
+
"""
|
794 |
+
def __init__(
|
795 |
+
self,
|
796 |
+
embedding_dim: int = 512,
|
797 |
+
num_pos: int = 1024) -> None:
|
798 |
+
super(LearnedAbsolutePositionEmbedding1D, self).__init__()
|
799 |
+
self.embeddings = nn.Embedding(num_pos, embedding_dim)
|
800 |
+
self.num_pos = num_pos
|
801 |
+
|
802 |
+
def forward(self, seq_embeds: torch.Tensor) -> torch.Tensor:
|
803 |
+
"""
|
804 |
+
Args:
|
805 |
+
seq_embeds: The sequence embeddings in order. Allowed size:
|
806 |
+
1. [T, D], where T is the length of the sequence, and D is the
|
807 |
+
frame embedding dimension.
|
808 |
+
2. [B, T, D], where B is the batch size and T and D are the
|
809 |
+
same as above.
|
810 |
+
Returns a tensor of with the same dimensions as the input: i.e.,
|
811 |
+
[1, T, D] or [T, D].
|
812 |
+
"""
|
813 |
+
shape_len = len(seq_embeds.shape)
|
814 |
+
assert 2 <= shape_len <= 3
|
815 |
+
len_seq = seq_embeds.size(-2)
|
816 |
+
assert len_seq <= self.num_pos
|
817 |
+
# [T, D]
|
818 |
+
pos_embeds = self.embeddings(torch.arange(len_seq).to(seq_embeds.device))
|
819 |
+
# Adapt pre-computed positional embeddings to the input.
|
820 |
+
if shape_len == 3:
|
821 |
+
pos_embeds = pos_embeds.view(
|
822 |
+
(1, pos_embeds.size(0), pos_embeds.size(1)))
|
823 |
+
return pos_embeds
|
824 |
+
|
825 |
+
def create_git_attention_mask(
|
826 |
+
tgt: torch.Tensor,
|
827 |
+
memory: torch.Tensor,
|
828 |
+
max_length: int
|
829 |
+
) -> torch.Tensor:
|
830 |
+
# Obtain the dimensions of the target text and memory
|
831 |
+
batch_size = tgt.size(0)
|
832 |
+
num_tgt = tgt.shape[1]
|
833 |
+
num_memory = memory.shape[1]
|
834 |
+
total_length = num_memory + num_tgt
|
835 |
+
|
836 |
+
# Create the top left part of the attention matrix
|
837 |
+
top_left = torch.zeros((num_memory, num_memory)) # Attention enabled in this region
|
838 |
+
top_right = torch.full((num_memory, num_tgt), float(-3.4028e+38)) # Attention disabled here
|
839 |
+
|
840 |
+
# Bottom left part of the attention matrix
|
841 |
+
bottom_left = torch.zeros((num_tgt, num_memory)) # Attention enabled here
|
842 |
+
|
843 |
+
# Create a lower triangular matrix for the bottom right part
|
844 |
+
bottom_right = torch.tril(torch.ones(num_tgt, num_tgt))
|
845 |
+
|
846 |
+
# Transform 1s to 0 to enable attention, and 0s to -inf to block attention
|
847 |
+
bottom_right = bottom_right.masked_fill(bottom_right == 0, float(-3.4028e+38))
|
848 |
+
bottom_right = bottom_right.masked_fill(bottom_right == 1, float(0))
|
849 |
+
|
850 |
+
# Concatenate matrices to form the full mask
|
851 |
+
left = torch.cat((top_left, bottom_left), dim=0)
|
852 |
+
right = torch.cat((top_right, bottom_right), dim=0)
|
853 |
+
|
854 |
+
# Combine left and right parts
|
855 |
+
full_attention_mask = torch.cat((left, right), dim=1)
|
856 |
+
|
857 |
+
# Add padding to reach max_length
|
858 |
+
padding = torch.full((total_length, max_length - total_length), float(-3.4028e+38))
|
859 |
+
full_attention_mask = torch.cat((full_attention_mask, padding), dim=1)
|
860 |
+
|
861 |
+
# Add an axis for multi-heads and batch_size
|
862 |
+
full_attention_mask = full_attention_mask[None, None, :, :]
|
863 |
+
|
864 |
+
# Expand the mask to have shape (batch_size, 1, seq_length, max_length)
|
865 |
+
full_attention_mask = full_attention_mask.expand(batch_size, 1, full_attention_mask.size(-2), full_attention_mask.size(-1))
|
866 |
+
|
867 |
+
return full_attention_mask
|
868 |
+
|
869 |
+
def get_position_ids_from_binary_attention_mask(mask):
|
870 |
+
"""
|
871 |
+
Extract position IDs from a binary attention mask.
|
872 |
+
|
873 |
+
Args:
|
874 |
+
mask (torch.Tensor): The attention mask tensor of shape (1, 1, seq_len, seq_len),
|
875 |
+
where 1 indicates allowed attention and 0 indicates blocked attention.
|
876 |
+
|
877 |
+
Returns:
|
878 |
+
list: A list of lists where each sublist contains the allowed position IDs for each query position.
|
879 |
+
"""
|
880 |
+
# Assuming the mask is of shape (1, 1, seq_len, seq_len)
|
881 |
+
_, _, seq_len, _ = mask.shape
|
882 |
+
|
883 |
+
# Create a tensor with position IDs from 0 to seq_len - 1
|
884 |
+
position_ids = torch.arange(seq_len, dtype=torch.long, device=mask.device)
|
885 |
+
|
886 |
+
# Add a batch dimension
|
887 |
+
position_ids = position_ids.unsqueeze(0)
|
888 |
+
|
889 |
+
return position_ids
|
890 |
+
|
891 |
+
def ensure_tensor(variable):
|
892 |
+
# Check if the variable is a torch.Tensor
|
893 |
+
if isinstance(variable, torch.Tensor):
|
894 |
+
# print("Variable is already a tensor.")
|
895 |
+
return variable
|
896 |
+
else:
|
897 |
+
#print("Variable is not a tensor, converting...")
|
898 |
+
try:
|
899 |
+
# Convert the variable to a tensor
|
900 |
+
tensor = torch.tensor(variable)
|
901 |
+
#print("Conversion successful.")
|
902 |
+
return tensor
|
903 |
+
except Exception as e:
|
904 |
+
print(f"Error converting to tensor: {e}")
|
905 |
+
raise
|
906 |
+
|
907 |
+
@add_start_docstrings(
|
908 |
+
"The bare Model outputting raw hidden-states without any specific head on top.",
|
909 |
+
FEYNMODEL_START_DOCSTRING,
|
910 |
+
)
|
911 |
+
class FeynModel(Gemma2Model):
|
912 |
+
"""
|
913 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers.
|
914 |
+
Each layer is a [`FeynModelDecoderLayer`] + ['LoraLayer'] for *proj* moduls
|
915 |
+
NB : LoraLayers will be added and activatd on proj modules onpy if pixel_values is not None
|
916 |
+
|
917 |
+
Args:
|
918 |
+
config: FeynModelConfig
|
919 |
+
"""
|
920 |
+
|
921 |
+
def __init__(self, config: FeynModelConfig):
|
922 |
+
super().__init__(config)
|
923 |
+
# Initialize weights and apply final processing
|
924 |
+
self.mode='llm'
|
925 |
+
'''
|
926 |
+
self.image_patch_tokens = int(
|
927 |
+
(config.vision_config.image_size / config.vision_config.patch_size) ** 2 + 1
|
928 |
+
)
|
929 |
+
|
930 |
+
if config.num_image_with_embedding is not None:
|
931 |
+
self.image_patch_tokens *= config.num_image_with_embedding
|
932 |
+
'''
|
933 |
+
self.image_patch_tokens = 577
|
934 |
+
self.post_init()
|
935 |
+
|
936 |
+
def get_input_embeddings(self):
|
937 |
+
return self.embed_tokens
|
938 |
+
|
939 |
+
def set_input_embeddings(self, value):
|
940 |
+
self.embed_tokens = value
|
941 |
+
|
942 |
+
|
943 |
+
|
944 |
+
|
945 |
+
@add_start_docstrings_to_model_forward(FEYNMODEL_INPUTS_DOCSTRING)
|
946 |
+
def forward(
|
947 |
+
self,
|
948 |
+
input_ids: torch.LongTensor = None,
|
949 |
+
attention_mask: Optional[torch.Tensor] = None,
|
950 |
+
position_ids: Optional[torch.LongTensor] = None,
|
951 |
+
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
952 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
953 |
+
use_cache: Optional[bool] = None,
|
954 |
+
output_attentions: Optional[bool] = None,
|
955 |
+
output_hidden_states: Optional[bool] = None,
|
956 |
+
return_dict: Optional[bool] = None,
|
957 |
+
cache_position: Optional[torch.LongTensor] = None,
|
958 |
+
causal_attention_mask: Optional[torch.Tensor] = None,
|
959 |
+
**kwargs,
|
960 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
961 |
+
|
962 |
+
# print(f" self.mode = {self.mode}")
|
963 |
+
# Ensure cache_position is initialized if not provided
|
964 |
+
|
965 |
+
|
966 |
+
if cache_position is None:
|
967 |
+
batch_size = input_ids.size(0) if input_ids is not None else inputs_embeds.size(0)
|
968 |
+
cache_position = torch.zeros((batch_size,), dtype=torch.long, device=input_ids.device if input_ids is not None else inputs_embeds.device)
|
969 |
+
|
970 |
+
|
971 |
+
|
972 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
973 |
+
output_hidden_states = (
|
974 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
975 |
+
)
|
976 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
977 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
978 |
+
|
979 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
980 |
+
raise ValueError(
|
981 |
+
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
|
982 |
+
)
|
983 |
+
|
984 |
+
if self.gradient_checkpointing and self.training and use_cache:
|
985 |
+
logger.warning_once(
|
986 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
987 |
+
)
|
988 |
+
use_cache = False
|
989 |
+
|
990 |
+
if inputs_embeds is None:
|
991 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
992 |
+
causal_mask = self._update_causal_mask(
|
993 |
+
attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
|
994 |
+
)
|
995 |
+
else:
|
996 |
+
causal_mask = ensure_tensor(causal_attention_mask)
|
997 |
+
position_ids = get_position_ids_from_binary_attention_mask(attention_mask)
|
998 |
+
|
999 |
+
#print(f" causal_mask = {causal_mask} ")
|
1000 |
+
|
1001 |
+
if cache_position is None:
|
1002 |
+
cache_position = torch.arange(0, inputs_embeds.shape[1], device=inputs_embeds.device)
|
1003 |
+
|
1004 |
+
if position_ids is None :
|
1005 |
+
position_ids = cache_position.unsqueeze(0)
|
1006 |
+
|
1007 |
+
|
1008 |
+
|
1009 |
+
# Convert position_ids to a tensor if not already
|
1010 |
+
if not isinstance(position_ids, torch.Tensor):
|
1011 |
+
|
1012 |
+
position_ids = torch.tensor(position_ids, dtype=torch.long, device=inputs_embeds.device)
|
1013 |
+
|
1014 |
+
|
1015 |
+
# embed positions
|
1016 |
+
hidden_states = inputs_embeds
|
1017 |
+
|
1018 |
+
# normalized
|
1019 |
+
# FeynModel downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
|
1020 |
+
# See https://github.com/huggingface/transformers/pull/29402
|
1021 |
+
normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
|
1022 |
+
hidden_states = hidden_states * normalizer
|
1023 |
+
|
1024 |
+
all_hidden_states = () if output_hidden_states else None
|
1025 |
+
all_self_attns = () if output_attentions else None
|
1026 |
+
|
1027 |
+
for decoder_layer in self.layers:
|
1028 |
+
if output_hidden_states:
|
1029 |
+
all_hidden_states += (hidden_states,)
|
1030 |
+
|
1031 |
+
if self.gradient_checkpointing and self.training:
|
1032 |
+
layer_outputs = self._gradient_checkpointing_func(
|
1033 |
+
decoder_layer.__call__,
|
1034 |
+
hidden_states,
|
1035 |
+
causal_mask,
|
1036 |
+
position_ids,
|
1037 |
+
past_key_values,
|
1038 |
+
output_attentions,
|
1039 |
+
use_cache,
|
1040 |
+
cache_position,
|
1041 |
+
)
|
1042 |
+
else:
|
1043 |
+
layer_outputs = decoder_layer(
|
1044 |
+
hidden_states,
|
1045 |
+
attention_mask=causal_mask,
|
1046 |
+
position_ids=position_ids,
|
1047 |
+
past_key_value=past_key_values,
|
1048 |
+
output_attentions=output_attentions,
|
1049 |
+
use_cache=use_cache,
|
1050 |
+
cache_position=cache_position,
|
1051 |
+
)
|
1052 |
+
|
1053 |
+
hidden_states = layer_outputs[0]
|
1054 |
+
|
1055 |
+
if output_attentions:
|
1056 |
+
all_self_attns += (layer_outputs[1],)
|
1057 |
+
|
1058 |
+
hidden_states = self.norm(hidden_states)
|
1059 |
+
|
1060 |
+
# add hidden states from the last decoder layer
|
1061 |
+
if output_hidden_states:
|
1062 |
+
all_hidden_states += (hidden_states,)
|
1063 |
+
|
1064 |
+
next_cache = past_key_values if use_cache else None
|
1065 |
+
|
1066 |
+
if not return_dict:
|
1067 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
1068 |
+
return BaseModelOutputWithPast(
|
1069 |
+
last_hidden_state=hidden_states,
|
1070 |
+
past_key_values=next_cache,
|
1071 |
+
hidden_states=all_hidden_states,
|
1072 |
+
attentions=all_self_attns,
|
1073 |
+
)
|
1074 |
+
|
1075 |
+
|
1076 |
+
|
1077 |
+
def _update_causal_mask(
|
1078 |
+
self,
|
1079 |
+
attention_mask: torch.Tensor,
|
1080 |
+
input_tensor: torch.Tensor,
|
1081 |
+
cache_position: torch.Tensor,
|
1082 |
+
past_key_values: Cache,
|
1083 |
+
output_attentions: bool,
|
1084 |
+
):
|
1085 |
+
|
1086 |
+
# print(f" _start _____ _update_causal_mask attention_mask {attention_mask.size()} {attention_mask} ")
|
1087 |
+
# Flash Attention currently doesn't support static cache but FeynModel work only with static cache.
|
1088 |
+
# So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape
|
1089 |
+
# to cut out keys/values trailing 0 used in static cache. This workaround should be compile compatible
|
1090 |
+
# as it doesn't cause dynamic control issues.
|
1091 |
+
if self.config._attn_implementation == "flash_attention_2":
|
1092 |
+
return attention_mask
|
1093 |
+
|
1094 |
+
dtype, device = input_tensor.dtype, input_tensor.device
|
1095 |
+
min_dtype = torch.finfo(dtype).min
|
1096 |
+
sequence_length = input_tensor.shape[1]
|
1097 |
+
if isinstance(past_key_values, HybridCache):
|
1098 |
+
target_length = past_key_values.get_max_length()
|
1099 |
+
else:
|
1100 |
+
target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]
|
1101 |
+
|
1102 |
+
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
1103 |
+
causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
|
1104 |
+
attention_mask,
|
1105 |
+
sequence_length=sequence_length,
|
1106 |
+
target_length=target_length,
|
1107 |
+
dtype=dtype,
|
1108 |
+
device=device,
|
1109 |
+
min_dtype=min_dtype,
|
1110 |
+
cache_position=cache_position,
|
1111 |
+
batch_size=input_tensor.shape[0],
|
1112 |
+
)
|
1113 |
+
#print(f" _end ______ _update_causal_mask causal_mask {causal_mask.size()} {causal_mask} ")
|
1114 |
+
return causal_mask
|
1115 |
+
|
1116 |
+
|
1117 |
+
|
1118 |
+
class FeynModelForCausalLM(Gemma2ForCausalLM):
|
1119 |
+
_tied_weights_keys = ["lm_head.weight"]
|
1120 |
+
config_class = FeynModelConfig
|
1121 |
+
def __init__(self, config):
|
1122 |
+
super().__init__(config)
|
1123 |
+
config.vision_config=Florence2VisionConfig.from_dict(config.vision_config)
|
1124 |
+
self.model = FeynModel(config)
|
1125 |
+
|
1126 |
+
# assert config.vision_config.model_type== 'davit', 'only DaViT is supported for now'
|
1127 |
+
self.vision_tower = DaViT.from_config(config=config.vision_config)
|
1128 |
+
self._build_image_projection_layers(config)
|
1129 |
+
|
1130 |
+
self.__causal_attention_mask = None
|
1131 |
+
|
1132 |
+
# Initialize weights and apply final processing
|
1133 |
+
self.post_init()
|
1134 |
+
|
1135 |
+
################ Vision Tower ########################
|
1136 |
+
def _build_image_projection_layers(self, config):
|
1137 |
+
image_dim_out = config.vision_config.dim_embed[-1]
|
1138 |
+
dim_projection = config.vision_config.projection_dim
|
1139 |
+
self.image_projection = nn.Parameter(
|
1140 |
+
torch.empty(image_dim_out, dim_projection)
|
1141 |
+
)
|
1142 |
+
self.image_proj_norm = nn.LayerNorm(dim_projection)
|
1143 |
+
image_pos_embed_config = config.vision_config.image_pos_embed
|
1144 |
+
if image_pos_embed_config['type'] == 'learned_abs_2d':
|
1145 |
+
self.image_pos_embed = LearnedAbsolutePositionEmbedding2D(
|
1146 |
+
embedding_dim=image_dim_out,
|
1147 |
+
num_pos=image_pos_embed_config['max_pos_embeddings']
|
1148 |
+
)
|
1149 |
+
else:
|
1150 |
+
raise NotImplementedError('Not implemented yet')
|
1151 |
+
|
1152 |
+
self.image_feature_source = config.vision_config.image_feature_source
|
1153 |
+
|
1154 |
+
# temporal embedding
|
1155 |
+
visual_temporal_embedding_config = config.vision_config.visual_temporal_embedding
|
1156 |
+
if visual_temporal_embedding_config['type'] == 'COSINE':
|
1157 |
+
self.visual_temporal_embed = PositionalEmbeddingCosine1D(
|
1158 |
+
embed_dim=image_dim_out,
|
1159 |
+
max_seq_len=visual_temporal_embedding_config['max_temporal_embeddings']
|
1160 |
+
)
|
1161 |
+
else:
|
1162 |
+
raise NotImplementedError('Not implemented yet')
|
1163 |
+
|
1164 |
+
|
1165 |
+
|
1166 |
+
def _merge_input_ids_with_image_features(self, image_features, inputs_embeds):
|
1167 |
+
batch_size, image_token_length = image_features.size()[:-1]
|
1168 |
+
device = image_features.device
|
1169 |
+
image_attention_mask = torch.ones(batch_size, image_token_length, device=device)
|
1170 |
+
|
1171 |
+
if inputs_embeds is None:
|
1172 |
+
return image_features, image_attention_mask
|
1173 |
+
|
1174 |
+
task_prefix_embeds = inputs_embeds
|
1175 |
+
task_prefix_attention_mask = torch.ones(batch_size, task_prefix_embeds.size(1), device=device)
|
1176 |
+
|
1177 |
+
# Assurer que les masques d'attention sont de deux dimensions
|
1178 |
+
if len(task_prefix_attention_mask.shape) == 3:
|
1179 |
+
task_prefix_attention_mask = task_prefix_attention_mask.squeeze(1)
|
1180 |
+
|
1181 |
+
# Vérifier la dimension de batch et ajuster si nécessaire
|
1182 |
+
if image_features.size(0) != task_prefix_embeds.size(0):
|
1183 |
+
raise ValueError("Batch sizes of image_features and task_prefix_embeds do not match")
|
1184 |
+
|
1185 |
+
# Ajouter une dimension fictive si les dimensions ne sont pas alignées
|
1186 |
+
if image_features.dim() < task_prefix_embeds.dim():
|
1187 |
+
image_features = image_features.unsqueeze(-1)
|
1188 |
+
elif task_prefix_embeds.dim() < image_features.dim():
|
1189 |
+
task_prefix_embeds = task_prefix_embeds.unsqueeze(-1)
|
1190 |
+
|
1191 |
+
# Assurer que toutes les dimensions, sauf dim=1, sont identiques
|
1192 |
+
if image_features.size(2) != task_prefix_embeds.size(2):
|
1193 |
+
# Ajuster ou signaler une erreur si les dimensions internes ne sont pas compatibles
|
1194 |
+
raise ValueError("Internal dimensions of image_features and task_prefix_embeds do not match")
|
1195 |
+
|
1196 |
+
inputs_embeds = torch.cat([image_features, task_prefix_embeds], dim=1)
|
1197 |
+
attention_mask = torch.cat([image_attention_mask, task_prefix_attention_mask], dim=1)
|
1198 |
+
|
1199 |
+
return inputs_embeds, attention_mask
|
1200 |
+
|
1201 |
+
def _encode_image(self, pixel_values):
|
1202 |
+
if len(pixel_values.shape) == 4:
|
1203 |
+
batch_size, C, H, W = pixel_values.shape
|
1204 |
+
T = 1
|
1205 |
+
x = self.vision_tower.forward_features_unpool(pixel_values)
|
1206 |
+
else:
|
1207 |
+
# Ajoute une dimension de batch au début si 'pixel_values' n'a que 3 dimensions (C, H, W)
|
1208 |
+
pixel_values = pixel_values.unsqueeze(0) # Ajoute une dimension de batch
|
1209 |
+
batch_size, C, H, W = pixel_values.shape
|
1210 |
+
T = 1
|
1211 |
+
x = self.vision_tower.forward_features_unpool(pixel_values)
|
1212 |
+
|
1213 |
+
if self.image_pos_embed is not None:
|
1214 |
+
x = x.view(batch_size * T, -1, x.shape[-1])
|
1215 |
+
num_tokens = x.shape[-2]
|
1216 |
+
h, w = int(num_tokens ** 0.5), int(num_tokens ** 0.5)
|
1217 |
+
assert h * w == num_tokens, 'only support square feature maps for now'
|
1218 |
+
x = x.view(batch_size * T, h, w, x.shape[-1])
|
1219 |
+
pos_embed = self.image_pos_embed(x)
|
1220 |
+
x = x + pos_embed
|
1221 |
+
x = x.view(batch_size, T * h*w, x.shape[-1])
|
1222 |
+
|
1223 |
+
if self.visual_temporal_embed is not None:
|
1224 |
+
visual_temporal_embed = self.visual_temporal_embed(x.view(batch_size, T, -1, x.shape[-1])[:, :, 0])
|
1225 |
+
x = x.view(batch_size, T, -1, x.shape[-1]) + visual_temporal_embed.view(1, T, 1, x.shape[-1])
|
1226 |
+
|
1227 |
+
x_feat_dict = {}
|
1228 |
+
|
1229 |
+
spatial_avg_pool_x = x.view(batch_size, T, -1, x.shape[-1]).mean(dim=2)
|
1230 |
+
x_feat_dict['spatial_avg_pool'] = spatial_avg_pool_x
|
1231 |
+
|
1232 |
+
temporal_avg_pool_x = x.view(batch_size, T, -1, x.shape[-1]).mean(dim=1)
|
1233 |
+
x_feat_dict['temporal_avg_pool'] = temporal_avg_pool_x
|
1234 |
+
|
1235 |
+
x = x.view(batch_size, T, -1, x.shape[-1])[:, -1]
|
1236 |
+
x_feat_dict['last_frame'] = x
|
1237 |
+
|
1238 |
+
new_x = []
|
1239 |
+
for _image_feature_source in self.image_feature_source:
|
1240 |
+
if _image_feature_source not in x_feat_dict:
|
1241 |
+
raise ValueError('invalid image feature source: {}'.format(_image_feature_source))
|
1242 |
+
new_x.append(x_feat_dict[_image_feature_source])
|
1243 |
+
|
1244 |
+
x = torch.cat(new_x, dim=1)
|
1245 |
+
|
1246 |
+
x = x @ self.image_projection
|
1247 |
+
x = self.image_proj_norm(x)
|
1248 |
+
|
1249 |
+
return x
|
1250 |
+
#######################################################
|
1251 |
+
|
1252 |
+
def get_input_embeddings(self):
|
1253 |
+
return self.model.embed_tokens
|
1254 |
+
|
1255 |
+
def set_input_embeddings(self, value):
|
1256 |
+
self.model.embed_tokens = value
|
1257 |
+
|
1258 |
+
def get_output_embeddings(self):
|
1259 |
+
return self.lm_head
|
1260 |
+
|
1261 |
+
def set_output_embeddings(self, new_embeddings):
|
1262 |
+
self.lm_head = new_embeddings
|
1263 |
+
|
1264 |
+
def set_decoder(self, decoder):
|
1265 |
+
self.model = decoder
|
1266 |
+
|
1267 |
+
def get_decoder(self):
|
1268 |
+
return self.model
|
1269 |
+
|
1270 |
+
@add_start_docstrings_to_model_forward(FEYNMODEL_INPUTS_DOCSTRING)
|
1271 |
+
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
1272 |
+
def forward(
|
1273 |
+
self,
|
1274 |
+
input_ids: torch.LongTensor = None,
|
1275 |
+
pixel_values: Optional[torch.Tensor] = None,
|
1276 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1277 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1278 |
+
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
1279 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1280 |
+
labels: Optional[torch.LongTensor] = None,
|
1281 |
+
use_cache: Optional[bool] = None,
|
1282 |
+
output_attentions: Optional[bool] = None,
|
1283 |
+
output_hidden_states: Optional[bool] = None,
|
1284 |
+
return_dict: Optional[bool] = None,
|
1285 |
+
cache_position: Optional[torch.LongTensor] = None,
|
1286 |
+
**kwargs,
|
1287 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
1288 |
+
r"""
|
1289 |
+
Args:
|
1290 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1291 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
1292 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
1293 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
1294 |
+
|
1295 |
+
Returns:
|
1296 |
+
|
1297 |
+
Example:
|
1298 |
+
|
1299 |
+
```python
|
1300 |
+
>>> from transformers import AutoTokenizer, GemmaForCausalLM
|
1301 |
+
|
1302 |
+
>>> model = GemmaForCausalLM.from_pretrained("google/gemma-2-9b")
|
1303 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
|
1304 |
+
|
1305 |
+
>>> prompt = "What is your favorite condiment?"
|
1306 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
1307 |
+
|
1308 |
+
>>> # Generate
|
1309 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
1310 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
1311 |
+
"What is your favorite condiment?"
|
1312 |
+
```"""
|
1313 |
+
|
1314 |
+
|
1315 |
+
if self.training and self.config._attn_implementation != "eager":
|
1316 |
+
logger.warning_once(
|
1317 |
+
"It is strongly recommended to train FeynModel models with the `eager` attention implementation "
|
1318 |
+
f"instead of `{self.config._attn_implementation}`. Use `eager` with `AutoModelForCausalLM.from_pretrained('<path-to-checkpoint>', attn_implementation='eager')`."
|
1319 |
+
)
|
1320 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1321 |
+
output_hidden_states = (
|
1322 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1323 |
+
)
|
1324 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1325 |
+
|
1326 |
+
if pixel_values is not None:
|
1327 |
+
self.model.mode='vlm'
|
1328 |
+
|
1329 |
+
if input_ids is not None:
|
1330 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
1331 |
+
image_features = self._encode_image(pixel_values)
|
1332 |
+
inputs_embeds, causal_attention_mask = self._merge_input_ids_with_image_features(image_features, inputs_embeds )
|
1333 |
+
causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=2048)
|
1334 |
+
causal_attention_mask=causal_attention_mask.to(input_ids.device)
|
1335 |
+
self.__causal_attention_mask=causal_attention_mask
|
1336 |
+
|
1337 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
1338 |
+
if pixel_values is not None:
|
1339 |
+
outputs = self.model(
|
1340 |
+
input_ids=None,
|
1341 |
+
attention_mask=causal_attention_mask,
|
1342 |
+
position_ids=position_ids,
|
1343 |
+
past_key_values=past_key_values,
|
1344 |
+
inputs_embeds=inputs_embeds,
|
1345 |
+
use_cache=use_cache,
|
1346 |
+
output_attentions=output_attentions,
|
1347 |
+
output_hidden_states=output_hidden_states,
|
1348 |
+
return_dict=return_dict,
|
1349 |
+
cache_position=cache_position,
|
1350 |
+
causal_attention_mask=causal_attention_mask,
|
1351 |
+
)
|
1352 |
+
else:
|
1353 |
+
outputs = self.model(
|
1354 |
+
input_ids=input_ids,
|
1355 |
+
attention_mask=attention_mask,
|
1356 |
+
position_ids=position_ids,
|
1357 |
+
past_key_values=past_key_values,
|
1358 |
+
inputs_embeds=inputs_embeds,
|
1359 |
+
use_cache=use_cache,
|
1360 |
+
output_attentions=output_attentions,
|
1361 |
+
output_hidden_states=output_hidden_states,
|
1362 |
+
return_dict=return_dict,
|
1363 |
+
cache_position=cache_position,
|
1364 |
+
causal_attention_mask=self.__causal_attention_mask,
|
1365 |
+
)
|
1366 |
+
|
1367 |
+
|
1368 |
+
hidden_states = outputs[0]
|
1369 |
+
logits = self.lm_head(hidden_states)
|
1370 |
+
|
1371 |
+
if self.config.final_logit_softcapping is not None:
|
1372 |
+
logits = logits / self.config.final_logit_softcapping
|
1373 |
+
logits = torch.tanh(logits)
|
1374 |
+
logits = logits * self.config.final_logit_softcapping
|
1375 |
+
|
1376 |
+
|
1377 |
+
logits = logits.float()
|
1378 |
+
loss = None
|
1379 |
+
if labels is not None:
|
1380 |
+
# we are doing next-token prediction; shift prediction scores and input ids by one
|
1381 |
+
num_image_tokens = self.model.image_patch_tokens
|
1382 |
+
shifted_logits = logits[:, num_image_tokens:-1, :].contiguous()
|
1383 |
+
labels = labels[:, 1:].contiguous()
|
1384 |
+
loss_fct = CrossEntropyLoss()
|
1385 |
+
loss = loss_fct(shifted_logits.view(-1, self.config.vocab_size), labels.view(-1))
|
1386 |
+
|
1387 |
+
if not return_dict:
|
1388 |
+
|
1389 |
+
output = (logits,) + outputs[1:]
|
1390 |
+
return (loss,) + output if loss is not None else output
|
1391 |
+
|
1392 |
+
return CausalLMOutputWithPast(
|
1393 |
+
loss=loss,
|
1394 |
+
logits=logits,
|
1395 |
+
past_key_values=outputs.past_key_values,
|
1396 |
+
hidden_states=outputs.hidden_states,
|
1397 |
+
attentions=outputs.attentions,
|
1398 |
+
)
|
1399 |
+
|
1400 |
+
def prepare_inputs_for_generation(
|
1401 |
+
self,
|
1402 |
+
input_ids,
|
1403 |
+
past_key_values=None,
|
1404 |
+
attention_mask=None,
|
1405 |
+
inputs_embeds=None,
|
1406 |
+
cache_position=None,
|
1407 |
+
position_ids=None,
|
1408 |
+
use_cache=True,
|
1409 |
+
**kwargs,
|
1410 |
+
):
|
1411 |
+
|
1412 |
+
|
1413 |
+
# If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
|
1414 |
+
# Exception 1: when passing input_embeds, input_ids may be missing entries
|
1415 |
+
# Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
|
1416 |
+
if past_key_values is not None:
|
1417 |
+
if inputs_embeds is not None: # Exception 1
|
1418 |
+
input_ids = input_ids[:, -cache_position.shape[0] :]
|
1419 |
+
elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2)
|
1420 |
+
input_ids = input_ids[:, cache_position]
|
1421 |
+
|
1422 |
+
if attention_mask is not None and position_ids is None:
|
1423 |
+
# create position_ids on the fly for batch generation
|
1424 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
1425 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
1426 |
+
if past_key_values:
|
1427 |
+
# print(f"+-+-+-+-+-+-+++ past_key_values +-+-+++- position_ids {position_ids.size()} ================= ")
|
1428 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
1429 |
+
# This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s
|
1430 |
+
# `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride
|
1431 |
+
# during the decoding. Here, simply using `.contiguous()` is not sufficient as in the
|
1432 |
+
# batch size = 1 case, `position_ids` is already contiguous but with varying stride
|
1433 |
+
# which retriggers a capture.
|
1434 |
+
position_ids = position_ids.clone(memory_format=torch.contiguous_format)
|
1435 |
+
# print(f"+-+-+-+-+-+-+++ past_key_values +-+-+++- position_ids cmlone ==> {position_ids.size()} ================= ")
|
1436 |
+
|
1437 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
1438 |
+
if inputs_embeds is not None and cache_position[0] == 0:
|
1439 |
+
#print(">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> first generation step>>>>>>>>>>>>>>>>>>>>>>>>>>>>>><")
|
1440 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
1441 |
+
else:
|
1442 |
+
# The clone here is for the same reason as for `position_ids`.
|
1443 |
+
# print(">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> The clone here is for the same reason as for `position_ids` ==> input_ids input_ids.clone.>>>>>>>>>>>>>>>>>>>>>>>>>>>>>><")
|
1444 |
+
model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format)}
|
1445 |
+
|
1446 |
+
if isinstance(past_key_values, HybridCache) and attention_mask.ndim == 2:
|
1447 |
+
if inputs_embeds is not None and input_ids.size(1)!= 0 :
|
1448 |
+
###################### V ############## add _ for _ = inputs_embeds.shape
|
1449 |
+
batch_size, sequence_length, _ = inputs_embeds.shape
|
1450 |
+
device = inputs_embeds.device
|
1451 |
+
#print(f"1111111 +-+-+-+-+-+-+-+-+-+- sequence_length = inputs_embeds {sequence_length}")
|
1452 |
+
else:
|
1453 |
+
batch_size, sequence_length = position_ids.shape
|
1454 |
+
device = input_ids.device
|
1455 |
+
#print(f"22222222 +-+-+-+-+-+-+-+-+-+- sequence_length = input_ids.shape {sequence_length}")
|
1456 |
+
|
1457 |
+
dtype = self.lm_head.weight.dtype
|
1458 |
+
min_dtype = torch.finfo(dtype).min
|
1459 |
+
|
1460 |
+
attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
|
1461 |
+
attention_mask,
|
1462 |
+
sequence_length=sequence_length,
|
1463 |
+
target_length=past_key_values.get_max_length(),
|
1464 |
+
dtype=dtype,
|
1465 |
+
device=device,
|
1466 |
+
min_dtype=min_dtype,
|
1467 |
+
cache_position=cache_position,
|
1468 |
+
batch_size=batch_size,
|
1469 |
+
)
|
1470 |
+
|
1471 |
+
|
1472 |
+
model_inputs.update(
|
1473 |
+
{
|
1474 |
+
"position_ids": position_ids,
|
1475 |
+
"cache_position": cache_position,
|
1476 |
+
"past_key_values": past_key_values,
|
1477 |
+
"use_cache": use_cache,
|
1478 |
+
"attention_mask": attention_mask,
|
1479 |
+
}
|
1480 |
+
)
|
1481 |
+
return model_inputs
|
1482 |
+
|
1483 |
+
def generate(
|
1484 |
+
self,
|
1485 |
+
input_ids,
|
1486 |
+
pixel_values=None,
|
1487 |
+
max_length=None,
|
1488 |
+
do_sample=True,
|
1489 |
+
temperature=0.7,
|
1490 |
+
**kwargs
|
1491 |
+
):
|
1492 |
+
print("Fonction generate personnalisée appelée")
|
1493 |
+
|
1494 |
+
if pixel_values is not None:
|
1495 |
+
if input_ids is not None:
|
1496 |
+
print("input")
|
1497 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
1498 |
+
print("pixels")
|
1499 |
+
image_features = self._encode_image(pixel_values)
|
1500 |
+
inputs_embeds, causal_attention_mask = self._merge_input_ids_with_image_features(image_features, inputs_embeds )
|
1501 |
+
causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=max_length)
|
1502 |
+
causal_attention_mask=causal_attention_mask.to(input_ids.device)
|
1503 |
+
self.__causal_attention_mask=causal_attention_mask
|
1504 |
+
self.model.mode='vlm'
|
1505 |
+
result = super().generate(
|
1506 |
+
input_ids=None,
|
1507 |
+
inputs_embeds=inputs_embeds,
|
1508 |
+
max_length=max_length,
|
1509 |
+
do_sample=do_sample,
|
1510 |
+
temperature=temperature,
|
1511 |
+
**kwargs
|
1512 |
+
)
|
1513 |
+
|
1514 |
+
else:
|
1515 |
+
print("llm")
|
1516 |
+
self.model.mode=='llm'
|
1517 |
+
result = super().generate(
|
1518 |
+
input_ids=input_ids,
|
1519 |
+
#inputs_embeds=None,
|
1520 |
+
max_length=max_length,
|
1521 |
+
do_sample=do_sample,
|
1522 |
+
temperature=temperature,
|
1523 |
+
**kwargs
|
1524 |
+
)
|
1525 |
+
self.__causal_attention_mask = None
|
1526 |
+
|
1527 |
+
return result
|
1528 |
+
|
__init__.py
ADDED
File without changes
|
config.json
ADDED
@@ -0,0 +1,253 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"FeynModelForCausalLM"
|
4 |
+
],
|
5 |
+
"attention_bias": false,
|
6 |
+
"attention_dropout": 0.0,
|
7 |
+
"attn_logit_softcapping": 50.0,
|
8 |
+
"auto_map": {
|
9 |
+
"AutoConfig": "configuration_feynmodel.FeynModelConfig",
|
10 |
+
"AutoModelForCausalLM": "modeling_feynmodel.FeynModelForCausalLM"
|
11 |
+
},
|
12 |
+
"cache_implementation": "hybrid",
|
13 |
+
"final_logit_softcapping": 30.0,
|
14 |
+
"head_dim": 256,
|
15 |
+
"hidden_act": "gelu_pytorch_tanh",
|
16 |
+
"hidden_activation": "gelu_pytorch_tanh",
|
17 |
+
"hidden_size": 2304,
|
18 |
+
"ignore_index": -100,
|
19 |
+
"init_std": 0.02,
|
20 |
+
"initializer_range": 0.02,
|
21 |
+
"intermediate_size": 9216,
|
22 |
+
"max_position_embeddings": 8192,
|
23 |
+
"model_type": "FeynModel",
|
24 |
+
"num_attention_heads": 8,
|
25 |
+
"num_hidden_layers": 26,
|
26 |
+
"num_key_value_heads": 4,
|
27 |
+
"projection_dim": 1024,
|
28 |
+
"query_pre_attn_scalar": 256,
|
29 |
+
"rms_norm_eps": 1e-06,
|
30 |
+
"rope_theta": 10000.0,
|
31 |
+
"sliding_window": 4096,
|
32 |
+
"text_config": {
|
33 |
+
"_name_or_path": "Imagroune/feynmodel",
|
34 |
+
"add_cross_attention": false,
|
35 |
+
"architectures": [
|
36 |
+
"FeynModelForCausalLM"
|
37 |
+
],
|
38 |
+
"attention_bias": false,
|
39 |
+
"attention_dropout": 0.0,
|
40 |
+
"attn_logit_softcapping": 50.0,
|
41 |
+
"bad_words_ids": null,
|
42 |
+
"begin_suppress_tokens": null,
|
43 |
+
"bos_token_id": 2,
|
44 |
+
"cache_implementation": "hybrid",
|
45 |
+
"chunk_size_feed_forward": 0,
|
46 |
+
"cross_attention_hidden_size": null,
|
47 |
+
"decoder_start_token_id": null,
|
48 |
+
"diversity_penalty": 0.0,
|
49 |
+
"do_sample": false,
|
50 |
+
"early_stopping": false,
|
51 |
+
"encoder_no_repeat_ngram_size": 0,
|
52 |
+
"eos_token_id": [
|
53 |
+
1,
|
54 |
+
107
|
55 |
+
],
|
56 |
+
"exponential_decay_length_penalty": null,
|
57 |
+
"final_logit_softcapping": 30.0,
|
58 |
+
"finetuning_task": null,
|
59 |
+
"forced_bos_token_id": null,
|
60 |
+
"forced_eos_token_id": null,
|
61 |
+
"head_dim": 256,
|
62 |
+
"hidden_act": "gelu_pytorch_tanh",
|
63 |
+
"hidden_activation": "gelu_pytorch_tanh",
|
64 |
+
"hidden_size": 2304,
|
65 |
+
"id2label": {
|
66 |
+
"0": "LABEL_0",
|
67 |
+
"1": "LABEL_1"
|
68 |
+
},
|
69 |
+
"init_std": 0.02,
|
70 |
+
"initializer_range": 0.02,
|
71 |
+
"intermediate_size": 9216,
|
72 |
+
"is_decoder": false,
|
73 |
+
"is_encoder_decoder": false,
|
74 |
+
"label2id": {
|
75 |
+
"LABEL_0": 0,
|
76 |
+
"LABEL_1": 1
|
77 |
+
},
|
78 |
+
"length_penalty": 1.0,
|
79 |
+
"max_length": 20,
|
80 |
+
"max_position_embeddings": 8192,
|
81 |
+
"min_length": 0,
|
82 |
+
"model_type": "FeynModel",
|
83 |
+
"no_repeat_ngram_size": 0,
|
84 |
+
"num_attention_heads": 8,
|
85 |
+
"num_beam_groups": 1,
|
86 |
+
"num_beams": 1,
|
87 |
+
"num_hidden_layers": 26,
|
88 |
+
"num_key_value_heads": 4,
|
89 |
+
"num_return_sequences": 1,
|
90 |
+
"output_attentions": false,
|
91 |
+
"output_hidden_states": false,
|
92 |
+
"output_scores": false,
|
93 |
+
"pad_token_id": 0,
|
94 |
+
"prefix": null,
|
95 |
+
"problem_type": null,
|
96 |
+
"pruned_heads": {},
|
97 |
+
"query_pre_attn_scalar": 256,
|
98 |
+
"remove_invalid_values": false,
|
99 |
+
"repetition_penalty": 1.0,
|
100 |
+
"return_dict": true,
|
101 |
+
"return_dict_in_generate": false,
|
102 |
+
"rms_norm_eps": 1e-06,
|
103 |
+
"rope_theta": 10000.0,
|
104 |
+
"sep_token_id": null,
|
105 |
+
"sliding_window": 4096,
|
106 |
+
"suppress_tokens": null,
|
107 |
+
"task_specific_params": null,
|
108 |
+
"temperature": 1.0,
|
109 |
+
"tf_legacy_loss": false,
|
110 |
+
"tie_encoder_decoder": false,
|
111 |
+
"tie_word_embeddings": true,
|
112 |
+
"tokenizer_class": null,
|
113 |
+
"top_k": 50,
|
114 |
+
"top_p": 1.0,
|
115 |
+
"torch_dtype": "float16",
|
116 |
+
"torchscript": false,
|
117 |
+
"typical_p": 1.0,
|
118 |
+
"use_bfloat16": false,
|
119 |
+
"use_cache": true,
|
120 |
+
"vocab_size": 256000
|
121 |
+
},
|
122 |
+
"torch_dtype": "float32",
|
123 |
+
"transformers_version": "4.44.2",
|
124 |
+
"use_cache": true,
|
125 |
+
"vision_config": {
|
126 |
+
"_name_or_path": "",
|
127 |
+
"add_cross_attention": false,
|
128 |
+
"architectures": null,
|
129 |
+
"bad_words_ids": null,
|
130 |
+
"begin_suppress_tokens": null,
|
131 |
+
"bos_token_id": null,
|
132 |
+
"chunk_size_feed_forward": 0,
|
133 |
+
"cross_attention_hidden_size": null,
|
134 |
+
"decoder_start_token_id": null,
|
135 |
+
"depths": [
|
136 |
+
1,
|
137 |
+
1,
|
138 |
+
9,
|
139 |
+
1
|
140 |
+
],
|
141 |
+
"dim_embed": [
|
142 |
+
128,
|
143 |
+
256,
|
144 |
+
512,
|
145 |
+
1024
|
146 |
+
],
|
147 |
+
"diversity_penalty": 0.0,
|
148 |
+
"do_sample": false,
|
149 |
+
"drop_path_rate": 0.1,
|
150 |
+
"early_stopping": false,
|
151 |
+
"enable_checkpoint": false,
|
152 |
+
"encoder_no_repeat_ngram_size": 0,
|
153 |
+
"eos_token_id": null,
|
154 |
+
"exponential_decay_length_penalty": null,
|
155 |
+
"finetuning_task": null,
|
156 |
+
"forced_bos_token_id": null,
|
157 |
+
"forced_eos_token_id": null,
|
158 |
+
"id2label": {
|
159 |
+
"0": "LABEL_0",
|
160 |
+
"1": "LABEL_1"
|
161 |
+
},
|
162 |
+
"image_feature_source": [
|
163 |
+
"spatial_avg_pool",
|
164 |
+
"temporal_avg_pool"
|
165 |
+
],
|
166 |
+
"image_pos_embed": {
|
167 |
+
"max_pos_embeddings": 50,
|
168 |
+
"type": "learned_abs_2d"
|
169 |
+
},
|
170 |
+
"is_decoder": false,
|
171 |
+
"is_encoder_decoder": false,
|
172 |
+
"label2id": {
|
173 |
+
"LABEL_0": 0,
|
174 |
+
"LABEL_1": 1
|
175 |
+
},
|
176 |
+
"length_penalty": 1.0,
|
177 |
+
"max_length": 20,
|
178 |
+
"min_length": 0,
|
179 |
+
"model_type": "florence2_vision",
|
180 |
+
"no_repeat_ngram_size": 0,
|
181 |
+
"num_beam_groups": 1,
|
182 |
+
"num_beams": 1,
|
183 |
+
"num_groups": [
|
184 |
+
4,
|
185 |
+
8,
|
186 |
+
16,
|
187 |
+
32
|
188 |
+
],
|
189 |
+
"num_heads": [
|
190 |
+
4,
|
191 |
+
8,
|
192 |
+
16,
|
193 |
+
32
|
194 |
+
],
|
195 |
+
"num_return_sequences": 1,
|
196 |
+
"output_attentions": false,
|
197 |
+
"output_hidden_states": false,
|
198 |
+
"output_scores": false,
|
199 |
+
"pad_token_id": null,
|
200 |
+
"patch_padding": [
|
201 |
+
3,
|
202 |
+
1,
|
203 |
+
1,
|
204 |
+
1
|
205 |
+
],
|
206 |
+
"patch_prenorm": [
|
207 |
+
false,
|
208 |
+
true,
|
209 |
+
true,
|
210 |
+
true
|
211 |
+
],
|
212 |
+
"patch_size": [
|
213 |
+
7,
|
214 |
+
3,
|
215 |
+
3,
|
216 |
+
3
|
217 |
+
],
|
218 |
+
"patch_stride": [
|
219 |
+
4,
|
220 |
+
2,
|
221 |
+
2,
|
222 |
+
2
|
223 |
+
],
|
224 |
+
"prefix": null,
|
225 |
+
"problem_type": null,
|
226 |
+
"projection_dim": 2304,
|
227 |
+
"pruned_heads": {},
|
228 |
+
"remove_invalid_values": false,
|
229 |
+
"repetition_penalty": 1.0,
|
230 |
+
"return_dict": true,
|
231 |
+
"return_dict_in_generate": false,
|
232 |
+
"sep_token_id": null,
|
233 |
+
"suppress_tokens": null,
|
234 |
+
"task_specific_params": null,
|
235 |
+
"temperature": 1.0,
|
236 |
+
"tf_legacy_loss": false,
|
237 |
+
"tie_encoder_decoder": false,
|
238 |
+
"tie_word_embeddings": true,
|
239 |
+
"tokenizer_class": null,
|
240 |
+
"top_k": 50,
|
241 |
+
"top_p": 1.0,
|
242 |
+
"torch_dtype": null,
|
243 |
+
"torchscript": false,
|
244 |
+
"typical_p": 1.0,
|
245 |
+
"use_bfloat16": false,
|
246 |
+
"visual_temporal_embedding": {
|
247 |
+
"max_temporal_embeddings": 100,
|
248 |
+
"type": "COSINE"
|
249 |
+
},
|
250 |
+
"window_size": 12
|
251 |
+
},
|
252 |
+
"vocab_size": 256000
|
253 |
+
}
|
configuration_feynmodel.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig
|
2 |
+
import copy
|
3 |
+
|
4 |
+
class Florence2VisionConfig(PretrainedConfig):
|
5 |
+
|
6 |
+
model_type = "florence2_vision"
|
7 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
8 |
+
|
9 |
+
def __init__(
|
10 |
+
self,
|
11 |
+
drop_path_rate=0.1,
|
12 |
+
patch_size=[7, 3, 3, 3],
|
13 |
+
patch_stride=[4, 2, 2, 2],
|
14 |
+
patch_padding=[3, 1, 1, 1],
|
15 |
+
patch_prenorm=[False, True, True, True],
|
16 |
+
enable_checkpoint=False,
|
17 |
+
dim_embed=[256, 512, 1024, 2048],
|
18 |
+
num_heads=[8, 16, 32, 64],
|
19 |
+
num_groups=[8, 16, 32, 64],
|
20 |
+
depths=[1, 1, 9, 1],
|
21 |
+
window_size=12,
|
22 |
+
projection_dim=1024,
|
23 |
+
visual_temporal_embedding=None,
|
24 |
+
image_pos_embed=None,
|
25 |
+
image_feature_source=["spatial_avg_pool", "temporal_avg_pool"],
|
26 |
+
**kwargs,
|
27 |
+
):
|
28 |
+
self.drop_path_rate = drop_path_rate
|
29 |
+
self.patch_size = patch_size
|
30 |
+
self.patch_stride = patch_stride
|
31 |
+
self.patch_padding = patch_padding
|
32 |
+
self.patch_prenorm = patch_prenorm
|
33 |
+
self.enable_checkpoint = enable_checkpoint
|
34 |
+
self.dim_embed = dim_embed
|
35 |
+
self.num_heads = num_heads
|
36 |
+
self.num_groups = num_groups
|
37 |
+
self.depths = depths
|
38 |
+
self.window_size = window_size
|
39 |
+
self.projection_dim = projection_dim
|
40 |
+
self.visual_temporal_embedding = visual_temporal_embedding
|
41 |
+
self.image_pos_embed = image_pos_embed
|
42 |
+
self.image_feature_source = image_feature_source
|
43 |
+
|
44 |
+
super().__init__(**kwargs)
|
45 |
+
|
46 |
+
|
47 |
+
|
48 |
+
class Gemma2Config(PretrainedConfig):
|
49 |
+
|
50 |
+
model_type = "gemma2"
|
51 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
52 |
+
|
53 |
+
def __init__(
|
54 |
+
self,
|
55 |
+
vocab_size=256000,
|
56 |
+
hidden_size=3072,
|
57 |
+
intermediate_size=24576,
|
58 |
+
num_hidden_layers=28,
|
59 |
+
num_attention_heads=16,
|
60 |
+
num_key_value_heads=16,
|
61 |
+
head_dim=256,
|
62 |
+
hidden_activation="gelu_pytorch_tanh",
|
63 |
+
max_position_embeddings=8192,
|
64 |
+
initializer_range=0.02,
|
65 |
+
rms_norm_eps=1e-6,
|
66 |
+
use_cache=True,
|
67 |
+
pad_token_id=0,
|
68 |
+
eos_token_id=1,
|
69 |
+
bos_token_id=2,
|
70 |
+
tie_word_embeddings=True,
|
71 |
+
rope_theta=10000.0,
|
72 |
+
attention_bias=False,
|
73 |
+
attention_dropout=0.0,
|
74 |
+
final_logit_softcapping=30.0,
|
75 |
+
attn_logit_softcapping=50.0,
|
76 |
+
query_pre_attn_scalar=224,
|
77 |
+
sliding_window=4096,
|
78 |
+
**kwargs,
|
79 |
+
):
|
80 |
+
self.vocab_size = vocab_size
|
81 |
+
self.max_position_embeddings = max_position_embeddings
|
82 |
+
self.hidden_size = hidden_size
|
83 |
+
self.intermediate_size = intermediate_size
|
84 |
+
self.num_hidden_layers = num_hidden_layers
|
85 |
+
self.num_attention_heads = num_attention_heads
|
86 |
+
self.head_dim = head_dim
|
87 |
+
self.num_key_value_heads = num_key_value_heads
|
88 |
+
self.hidden_activation = hidden_activation
|
89 |
+
self.initializer_range = initializer_range
|
90 |
+
self.rms_norm_eps = rms_norm_eps
|
91 |
+
self.use_cache = use_cache
|
92 |
+
self.rope_theta = rope_theta
|
93 |
+
self.attention_bias = attention_bias
|
94 |
+
self.attention_dropout = attention_dropout
|
95 |
+
self.attn_logit_softcapping = attn_logit_softcapping
|
96 |
+
|
97 |
+
super().__init__(
|
98 |
+
pad_token_id=pad_token_id,
|
99 |
+
bos_token_id=bos_token_id,
|
100 |
+
eos_token_id=eos_token_id,
|
101 |
+
tie_word_embeddings=tie_word_embeddings,
|
102 |
+
**kwargs,
|
103 |
+
)
|
104 |
+
self.final_logit_softcapping = final_logit_softcapping
|
105 |
+
self.query_pre_attn_scalar = query_pre_attn_scalar
|
106 |
+
self.sliding_window = sliding_window
|
107 |
+
self.cache_implementation = "hybrid"
|
108 |
+
|
109 |
+
|
110 |
+
|
111 |
+
class FeynModelConfig(PretrainedConfig):
|
112 |
+
r"""
|
113 |
+
This is the configuration class to store the configuration of a [`FeynModel`]. It is used to instantiate a FeynModel
|
114 |
+
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
115 |
+
defaults will yield a similar configuration to that of the Gemma2-2B + Florence-2-Base + FeynModel V0.1.0.
|
116 |
+
```python
|
117 |
+
>>> from transformers import FeynModel, FeynModelConfig
|
118 |
+
>>> # Initializing a FeynModel style configuration
|
119 |
+
>>> configuration = FeynModelConfig()
|
120 |
+
>>> # Initializing a model
|
121 |
+
>>> model = FeynModel(configuration)
|
122 |
+
>>> # Accessing the model configuration
|
123 |
+
>>> configuration = model.config
|
124 |
+
```"""
|
125 |
+
|
126 |
+
# model_type = "gemma2"
|
127 |
+
# is_composition = False
|
128 |
+
model_type = "FeynModel"
|
129 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
130 |
+
|
131 |
+
def __init__(
|
132 |
+
self,
|
133 |
+
vision_config=None,
|
134 |
+
text_config=None,
|
135 |
+
ignore_index=-100,
|
136 |
+
vocab_size=256000,
|
137 |
+
projection_dim=1024,
|
138 |
+
**kwargs,
|
139 |
+
):
|
140 |
+
self.ignore_index = ignore_index
|
141 |
+
self.vocab_size = vocab_size
|
142 |
+
self.projection_dim = projection_dim
|
143 |
+
self.vision_config = vision_config
|
144 |
+
self.vocab_size = self.vocab_size
|
145 |
+
|
146 |
+
self.text_config = text_config
|
147 |
+
# self.sliding_window = text_config.sliding_window
|
148 |
+
# Ajout des attributs de text_config à l'instance actuelle de Config
|
149 |
+
|
150 |
+
if text_config is not None:
|
151 |
+
for attr, value in text_config.items():
|
152 |
+
setattr(self, attr, value)
|
153 |
+
|
154 |
+
|
155 |
+
super().__init__(**kwargs)
|
156 |
+
|
157 |
+
|
158 |
+
|
159 |
+
|
generation_config.json
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_from_model_config": true,
|
3 |
+
"bos_token_id": 2,
|
4 |
+
"cache_implementation": "hybrid",
|
5 |
+
"eos_token_id": [
|
6 |
+
1,
|
7 |
+
107
|
8 |
+
],
|
9 |
+
"pad_token_id": 0,
|
10 |
+
"transformers_version": "4.44.2"
|
11 |
+
}
|
model-00001-of-00003.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4214710d3c4b31d9a89527da2c196e344c9a41fbf6e4a7e942a8a626b9e911c5
|
3 |
+
size 4917078632
|
model-00002-of-00003.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e74788d2d95dc3174fd671a7e987fb4fb0243e25b1c8803a1fef8e084117638e
|
3 |
+
size 4983443424
|
model-00003-of-00003.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:24ca5c47d98179d7796291d5d4d6c2b8706f5173d2dcfc3fa57a6d394575f9fd
|
3 |
+
size 932581696
|
model.safetensors.index.json
ADDED
@@ -0,0 +1,705 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"metadata": {
|
3 |
+
"total_size": 10833017760
|
4 |
+
},
|
5 |
+
"weight_map": {
|
6 |
+
"image_pos_embed.column_embeddings.weight": "model-00003-of-00003.safetensors",
|
7 |
+
"image_pos_embed.row_embeddings.weight": "model-00003-of-00003.safetensors",
|
8 |
+
"image_proj_norm.bias": "model-00003-of-00003.safetensors",
|
9 |
+
"image_proj_norm.weight": "model-00003-of-00003.safetensors",
|
10 |
+
"image_projection": "model-00001-of-00003.safetensors",
|
11 |
+
"model.embed_tokens.weight": "model-00001-of-00003.safetensors",
|
12 |
+
"model.layers.0.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
13 |
+
"model.layers.0.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
14 |
+
"model.layers.0.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
15 |
+
"model.layers.0.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
16 |
+
"model.layers.0.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
17 |
+
"model.layers.0.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
18 |
+
"model.layers.0.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
19 |
+
"model.layers.0.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
20 |
+
"model.layers.0.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
21 |
+
"model.layers.0.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
22 |
+
"model.layers.0.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
23 |
+
"model.layers.1.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
24 |
+
"model.layers.1.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
25 |
+
"model.layers.1.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
26 |
+
"model.layers.1.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
27 |
+
"model.layers.1.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
28 |
+
"model.layers.1.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
29 |
+
"model.layers.1.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
30 |
+
"model.layers.1.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
31 |
+
"model.layers.1.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
32 |
+
"model.layers.1.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
33 |
+
"model.layers.1.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
34 |
+
"model.layers.10.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
35 |
+
"model.layers.10.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
36 |
+
"model.layers.10.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
37 |
+
"model.layers.10.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
38 |
+
"model.layers.10.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
39 |
+
"model.layers.10.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
40 |
+
"model.layers.10.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
41 |
+
"model.layers.10.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
42 |
+
"model.layers.10.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
43 |
+
"model.layers.10.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
44 |
+
"model.layers.10.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
45 |
+
"model.layers.11.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
46 |
+
"model.layers.11.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
47 |
+
"model.layers.11.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
48 |
+
"model.layers.11.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
49 |
+
"model.layers.11.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
50 |
+
"model.layers.11.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
51 |
+
"model.layers.11.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
52 |
+
"model.layers.11.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
53 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
54 |
+
"model.layers.11.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
55 |
+
"model.layers.11.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
56 |
+
"model.layers.12.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
57 |
+
"model.layers.12.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
58 |
+
"model.layers.12.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
59 |
+
"model.layers.12.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
60 |
+
"model.layers.12.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
61 |
+
"model.layers.12.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
62 |
+
"model.layers.12.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
63 |
+
"model.layers.12.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
64 |
+
"model.layers.12.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
65 |
+
"model.layers.12.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
66 |
+
"model.layers.12.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
67 |
+
"model.layers.13.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
68 |
+
"model.layers.13.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
69 |
+
"model.layers.13.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
70 |
+
"model.layers.13.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
71 |
+
"model.layers.13.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
72 |
+
"model.layers.13.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
73 |
+
"model.layers.13.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
74 |
+
"model.layers.13.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
75 |
+
"model.layers.13.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
76 |
+
"model.layers.13.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
77 |
+
"model.layers.13.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
78 |
+
"model.layers.14.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
79 |
+
"model.layers.14.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
80 |
+
"model.layers.14.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
81 |
+
"model.layers.14.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
82 |
+
"model.layers.14.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
83 |
+
"model.layers.14.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
84 |
+
"model.layers.14.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
85 |
+
"model.layers.14.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
86 |
+
"model.layers.14.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
87 |
+
"model.layers.14.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
88 |
+
"model.layers.14.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
89 |
+
"model.layers.15.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
90 |
+
"model.layers.15.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
91 |
+
"model.layers.15.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
92 |
+
"model.layers.15.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
93 |
+
"model.layers.15.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
94 |
+
"model.layers.15.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
95 |
+
"model.layers.15.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
96 |
+
"model.layers.15.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
97 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
98 |
+
"model.layers.15.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
99 |
+
"model.layers.15.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
100 |
+
"model.layers.16.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
101 |
+
"model.layers.16.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
102 |
+
"model.layers.16.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
103 |
+
"model.layers.16.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
104 |
+
"model.layers.16.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
105 |
+
"model.layers.16.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
106 |
+
"model.layers.16.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
107 |
+
"model.layers.16.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
108 |
+
"model.layers.16.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
109 |
+
"model.layers.16.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
110 |
+
"model.layers.16.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
111 |
+
"model.layers.17.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
112 |
+
"model.layers.17.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
113 |
+
"model.layers.17.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
114 |
+
"model.layers.17.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
115 |
+
"model.layers.17.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
116 |
+
"model.layers.17.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
117 |
+
"model.layers.17.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
118 |
+
"model.layers.17.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
119 |
+
"model.layers.17.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
120 |
+
"model.layers.17.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
121 |
+
"model.layers.17.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
122 |
+
"model.layers.18.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
123 |
+
"model.layers.18.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
124 |
+
"model.layers.18.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
125 |
+
"model.layers.18.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
126 |
+
"model.layers.18.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
127 |
+
"model.layers.18.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
128 |
+
"model.layers.18.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
129 |
+
"model.layers.18.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
130 |
+
"model.layers.18.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
131 |
+
"model.layers.18.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
132 |
+
"model.layers.18.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
133 |
+
"model.layers.19.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
134 |
+
"model.layers.19.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
135 |
+
"model.layers.19.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
136 |
+
"model.layers.19.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
137 |
+
"model.layers.19.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
138 |
+
"model.layers.19.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
139 |
+
"model.layers.19.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
140 |
+
"model.layers.19.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
141 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
142 |
+
"model.layers.19.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
143 |
+
"model.layers.19.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
144 |
+
"model.layers.2.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
145 |
+
"model.layers.2.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
146 |
+
"model.layers.2.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
147 |
+
"model.layers.2.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
148 |
+
"model.layers.2.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
149 |
+
"model.layers.2.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
150 |
+
"model.layers.2.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
151 |
+
"model.layers.2.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
152 |
+
"model.layers.2.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
153 |
+
"model.layers.2.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
154 |
+
"model.layers.2.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
155 |
+
"model.layers.20.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
156 |
+
"model.layers.20.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
157 |
+
"model.layers.20.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
158 |
+
"model.layers.20.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
159 |
+
"model.layers.20.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
160 |
+
"model.layers.20.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
161 |
+
"model.layers.20.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
162 |
+
"model.layers.20.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
163 |
+
"model.layers.20.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
164 |
+
"model.layers.20.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
165 |
+
"model.layers.20.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
166 |
+
"model.layers.21.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
167 |
+
"model.layers.21.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
168 |
+
"model.layers.21.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
169 |
+
"model.layers.21.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
170 |
+
"model.layers.21.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
171 |
+
"model.layers.21.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
172 |
+
"model.layers.21.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
173 |
+
"model.layers.21.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
174 |
+
"model.layers.21.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
175 |
+
"model.layers.21.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
176 |
+
"model.layers.21.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
177 |
+
"model.layers.22.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
178 |
+
"model.layers.22.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
179 |
+
"model.layers.22.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
180 |
+
"model.layers.22.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
181 |
+
"model.layers.22.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
182 |
+
"model.layers.22.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
183 |
+
"model.layers.22.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
184 |
+
"model.layers.22.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
185 |
+
"model.layers.22.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
186 |
+
"model.layers.22.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
187 |
+
"model.layers.22.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
188 |
+
"model.layers.23.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
189 |
+
"model.layers.23.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
190 |
+
"model.layers.23.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
191 |
+
"model.layers.23.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
192 |
+
"model.layers.23.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
193 |
+
"model.layers.23.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
194 |
+
"model.layers.23.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
195 |
+
"model.layers.23.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
196 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
197 |
+
"model.layers.23.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
198 |
+
"model.layers.23.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
199 |
+
"model.layers.24.input_layernorm.weight": "model-00003-of-00003.safetensors",
|
200 |
+
"model.layers.24.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
|
201 |
+
"model.layers.24.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
|
202 |
+
"model.layers.24.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
|
203 |
+
"model.layers.24.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
|
204 |
+
"model.layers.24.post_feedforward_layernorm.weight": "model-00003-of-00003.safetensors",
|
205 |
+
"model.layers.24.pre_feedforward_layernorm.weight": "model-00003-of-00003.safetensors",
|
206 |
+
"model.layers.24.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
207 |
+
"model.layers.24.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
208 |
+
"model.layers.24.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
209 |
+
"model.layers.24.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
210 |
+
"model.layers.25.input_layernorm.weight": "model-00003-of-00003.safetensors",
|
211 |
+
"model.layers.25.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
|
212 |
+
"model.layers.25.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
|
213 |
+
"model.layers.25.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
|
214 |
+
"model.layers.25.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
|
215 |
+
"model.layers.25.post_feedforward_layernorm.weight": "model-00003-of-00003.safetensors",
|
216 |
+
"model.layers.25.pre_feedforward_layernorm.weight": "model-00003-of-00003.safetensors",
|
217 |
+
"model.layers.25.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
|
218 |
+
"model.layers.25.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
|
219 |
+
"model.layers.25.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
|
220 |
+
"model.layers.25.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
|
221 |
+
"model.layers.3.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
222 |
+
"model.layers.3.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
223 |
+
"model.layers.3.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
224 |
+
"model.layers.3.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
225 |
+
"model.layers.3.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
226 |
+
"model.layers.3.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
227 |
+
"model.layers.3.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
228 |
+
"model.layers.3.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
229 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
230 |
+
"model.layers.3.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
231 |
+
"model.layers.3.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
232 |
+
"model.layers.4.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
233 |
+
"model.layers.4.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
234 |
+
"model.layers.4.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
235 |
+
"model.layers.4.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
236 |
+
"model.layers.4.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
237 |
+
"model.layers.4.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
238 |
+
"model.layers.4.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
239 |
+
"model.layers.4.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
240 |
+
"model.layers.4.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
241 |
+
"model.layers.4.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
242 |
+
"model.layers.4.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
243 |
+
"model.layers.5.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
244 |
+
"model.layers.5.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
245 |
+
"model.layers.5.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
246 |
+
"model.layers.5.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
247 |
+
"model.layers.5.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
248 |
+
"model.layers.5.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
249 |
+
"model.layers.5.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
250 |
+
"model.layers.5.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
251 |
+
"model.layers.5.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
252 |
+
"model.layers.5.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
253 |
+
"model.layers.5.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
254 |
+
"model.layers.6.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
255 |
+
"model.layers.6.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
256 |
+
"model.layers.6.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
257 |
+
"model.layers.6.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
258 |
+
"model.layers.6.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
259 |
+
"model.layers.6.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
260 |
+
"model.layers.6.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
261 |
+
"model.layers.6.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
262 |
+
"model.layers.6.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
263 |
+
"model.layers.6.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
264 |
+
"model.layers.6.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
265 |
+
"model.layers.7.input_layernorm.weight": "model-00001-of-00003.safetensors",
|
266 |
+
"model.layers.7.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
|
267 |
+
"model.layers.7.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
|
268 |
+
"model.layers.7.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
|
269 |
+
"model.layers.7.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
|
270 |
+
"model.layers.7.post_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
271 |
+
"model.layers.7.pre_feedforward_layernorm.weight": "model-00001-of-00003.safetensors",
|
272 |
+
"model.layers.7.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
273 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
274 |
+
"model.layers.7.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
275 |
+
"model.layers.7.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
276 |
+
"model.layers.8.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
277 |
+
"model.layers.8.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
278 |
+
"model.layers.8.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
279 |
+
"model.layers.8.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
280 |
+
"model.layers.8.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
281 |
+
"model.layers.8.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
282 |
+
"model.layers.8.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
283 |
+
"model.layers.8.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
|
284 |
+
"model.layers.8.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
|
285 |
+
"model.layers.8.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
|
286 |
+
"model.layers.8.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
|
287 |
+
"model.layers.9.input_layernorm.weight": "model-00002-of-00003.safetensors",
|
288 |
+
"model.layers.9.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
|
289 |
+
"model.layers.9.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
|
290 |
+
"model.layers.9.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
|
291 |
+
"model.layers.9.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
|
292 |
+
"model.layers.9.post_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
293 |
+
"model.layers.9.pre_feedforward_layernorm.weight": "model-00002-of-00003.safetensors",
|
294 |
+
"model.layers.9.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
|
295 |
+
"model.layers.9.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
|
296 |
+
"model.layers.9.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
|
297 |
+
"model.layers.9.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
|
298 |
+
"model.norm.weight": "model-00003-of-00003.safetensors",
|
299 |
+
"vision_tower.blocks.0.0.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
300 |
+
"vision_tower.blocks.0.0.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
301 |
+
"vision_tower.blocks.0.0.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
302 |
+
"vision_tower.blocks.0.0.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
303 |
+
"vision_tower.blocks.0.0.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
304 |
+
"vision_tower.blocks.0.0.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
305 |
+
"vision_tower.blocks.0.0.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
306 |
+
"vision_tower.blocks.0.0.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
307 |
+
"vision_tower.blocks.0.0.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
308 |
+
"vision_tower.blocks.0.0.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
309 |
+
"vision_tower.blocks.0.0.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
310 |
+
"vision_tower.blocks.0.0.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
311 |
+
"vision_tower.blocks.0.0.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
312 |
+
"vision_tower.blocks.0.0.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
313 |
+
"vision_tower.blocks.0.0.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
314 |
+
"vision_tower.blocks.0.0.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
315 |
+
"vision_tower.blocks.0.0.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
316 |
+
"vision_tower.blocks.0.0.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
317 |
+
"vision_tower.blocks.0.0.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
318 |
+
"vision_tower.blocks.0.0.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
319 |
+
"vision_tower.blocks.0.0.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
320 |
+
"vision_tower.blocks.0.0.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
321 |
+
"vision_tower.blocks.0.0.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
322 |
+
"vision_tower.blocks.0.0.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
323 |
+
"vision_tower.blocks.0.0.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
324 |
+
"vision_tower.blocks.0.0.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
325 |
+
"vision_tower.blocks.0.0.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
326 |
+
"vision_tower.blocks.0.0.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
327 |
+
"vision_tower.blocks.0.0.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
328 |
+
"vision_tower.blocks.0.0.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
329 |
+
"vision_tower.blocks.0.0.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
330 |
+
"vision_tower.blocks.0.0.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
331 |
+
"vision_tower.blocks.1.0.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
332 |
+
"vision_tower.blocks.1.0.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
333 |
+
"vision_tower.blocks.1.0.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
334 |
+
"vision_tower.blocks.1.0.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
335 |
+
"vision_tower.blocks.1.0.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
336 |
+
"vision_tower.blocks.1.0.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
337 |
+
"vision_tower.blocks.1.0.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
338 |
+
"vision_tower.blocks.1.0.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
339 |
+
"vision_tower.blocks.1.0.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
340 |
+
"vision_tower.blocks.1.0.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
341 |
+
"vision_tower.blocks.1.0.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
342 |
+
"vision_tower.blocks.1.0.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
343 |
+
"vision_tower.blocks.1.0.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
344 |
+
"vision_tower.blocks.1.0.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
345 |
+
"vision_tower.blocks.1.0.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
346 |
+
"vision_tower.blocks.1.0.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
347 |
+
"vision_tower.blocks.1.0.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
348 |
+
"vision_tower.blocks.1.0.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
349 |
+
"vision_tower.blocks.1.0.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
350 |
+
"vision_tower.blocks.1.0.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
351 |
+
"vision_tower.blocks.1.0.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
352 |
+
"vision_tower.blocks.1.0.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
353 |
+
"vision_tower.blocks.1.0.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
354 |
+
"vision_tower.blocks.1.0.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
355 |
+
"vision_tower.blocks.1.0.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
356 |
+
"vision_tower.blocks.1.0.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
357 |
+
"vision_tower.blocks.1.0.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
358 |
+
"vision_tower.blocks.1.0.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
359 |
+
"vision_tower.blocks.1.0.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
360 |
+
"vision_tower.blocks.1.0.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
361 |
+
"vision_tower.blocks.1.0.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
362 |
+
"vision_tower.blocks.1.0.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
363 |
+
"vision_tower.blocks.2.0.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
364 |
+
"vision_tower.blocks.2.0.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
365 |
+
"vision_tower.blocks.2.0.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
366 |
+
"vision_tower.blocks.2.0.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
367 |
+
"vision_tower.blocks.2.0.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
368 |
+
"vision_tower.blocks.2.0.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
369 |
+
"vision_tower.blocks.2.0.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
370 |
+
"vision_tower.blocks.2.0.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
371 |
+
"vision_tower.blocks.2.0.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
372 |
+
"vision_tower.blocks.2.0.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
373 |
+
"vision_tower.blocks.2.0.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
374 |
+
"vision_tower.blocks.2.0.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
375 |
+
"vision_tower.blocks.2.0.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
376 |
+
"vision_tower.blocks.2.0.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
377 |
+
"vision_tower.blocks.2.0.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
378 |
+
"vision_tower.blocks.2.0.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
379 |
+
"vision_tower.blocks.2.0.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
380 |
+
"vision_tower.blocks.2.0.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
381 |
+
"vision_tower.blocks.2.0.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
382 |
+
"vision_tower.blocks.2.0.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
383 |
+
"vision_tower.blocks.2.0.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
384 |
+
"vision_tower.blocks.2.0.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
385 |
+
"vision_tower.blocks.2.0.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
386 |
+
"vision_tower.blocks.2.0.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
387 |
+
"vision_tower.blocks.2.0.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
388 |
+
"vision_tower.blocks.2.0.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
389 |
+
"vision_tower.blocks.2.0.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
390 |
+
"vision_tower.blocks.2.0.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
391 |
+
"vision_tower.blocks.2.0.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
392 |
+
"vision_tower.blocks.2.0.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
393 |
+
"vision_tower.blocks.2.0.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
394 |
+
"vision_tower.blocks.2.0.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
395 |
+
"vision_tower.blocks.2.1.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
396 |
+
"vision_tower.blocks.2.1.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
397 |
+
"vision_tower.blocks.2.1.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
398 |
+
"vision_tower.blocks.2.1.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
399 |
+
"vision_tower.blocks.2.1.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
400 |
+
"vision_tower.blocks.2.1.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
401 |
+
"vision_tower.blocks.2.1.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
402 |
+
"vision_tower.blocks.2.1.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
403 |
+
"vision_tower.blocks.2.1.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
404 |
+
"vision_tower.blocks.2.1.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
405 |
+
"vision_tower.blocks.2.1.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
406 |
+
"vision_tower.blocks.2.1.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
407 |
+
"vision_tower.blocks.2.1.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
408 |
+
"vision_tower.blocks.2.1.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
409 |
+
"vision_tower.blocks.2.1.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
410 |
+
"vision_tower.blocks.2.1.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
411 |
+
"vision_tower.blocks.2.1.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
412 |
+
"vision_tower.blocks.2.1.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
413 |
+
"vision_tower.blocks.2.1.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
414 |
+
"vision_tower.blocks.2.1.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
415 |
+
"vision_tower.blocks.2.1.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
416 |
+
"vision_tower.blocks.2.1.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
417 |
+
"vision_tower.blocks.2.1.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
418 |
+
"vision_tower.blocks.2.1.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
419 |
+
"vision_tower.blocks.2.1.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
420 |
+
"vision_tower.blocks.2.1.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
421 |
+
"vision_tower.blocks.2.1.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
422 |
+
"vision_tower.blocks.2.1.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
423 |
+
"vision_tower.blocks.2.1.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
424 |
+
"vision_tower.blocks.2.1.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
425 |
+
"vision_tower.blocks.2.1.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
426 |
+
"vision_tower.blocks.2.1.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
427 |
+
"vision_tower.blocks.2.2.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
428 |
+
"vision_tower.blocks.2.2.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
429 |
+
"vision_tower.blocks.2.2.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
430 |
+
"vision_tower.blocks.2.2.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
431 |
+
"vision_tower.blocks.2.2.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
432 |
+
"vision_tower.blocks.2.2.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
433 |
+
"vision_tower.blocks.2.2.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
434 |
+
"vision_tower.blocks.2.2.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
435 |
+
"vision_tower.blocks.2.2.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
436 |
+
"vision_tower.blocks.2.2.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
437 |
+
"vision_tower.blocks.2.2.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
438 |
+
"vision_tower.blocks.2.2.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
439 |
+
"vision_tower.blocks.2.2.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
440 |
+
"vision_tower.blocks.2.2.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
441 |
+
"vision_tower.blocks.2.2.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
442 |
+
"vision_tower.blocks.2.2.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
443 |
+
"vision_tower.blocks.2.2.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
444 |
+
"vision_tower.blocks.2.2.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
445 |
+
"vision_tower.blocks.2.2.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
446 |
+
"vision_tower.blocks.2.2.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
447 |
+
"vision_tower.blocks.2.2.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
448 |
+
"vision_tower.blocks.2.2.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
449 |
+
"vision_tower.blocks.2.2.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
450 |
+
"vision_tower.blocks.2.2.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
451 |
+
"vision_tower.blocks.2.2.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
452 |
+
"vision_tower.blocks.2.2.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
453 |
+
"vision_tower.blocks.2.2.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
454 |
+
"vision_tower.blocks.2.2.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
455 |
+
"vision_tower.blocks.2.2.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
456 |
+
"vision_tower.blocks.2.2.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
457 |
+
"vision_tower.blocks.2.2.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
458 |
+
"vision_tower.blocks.2.2.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
459 |
+
"vision_tower.blocks.2.3.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
460 |
+
"vision_tower.blocks.2.3.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
461 |
+
"vision_tower.blocks.2.3.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
462 |
+
"vision_tower.blocks.2.3.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
463 |
+
"vision_tower.blocks.2.3.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
464 |
+
"vision_tower.blocks.2.3.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
465 |
+
"vision_tower.blocks.2.3.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
466 |
+
"vision_tower.blocks.2.3.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
467 |
+
"vision_tower.blocks.2.3.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
468 |
+
"vision_tower.blocks.2.3.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
469 |
+
"vision_tower.blocks.2.3.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
470 |
+
"vision_tower.blocks.2.3.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
471 |
+
"vision_tower.blocks.2.3.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
472 |
+
"vision_tower.blocks.2.3.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
473 |
+
"vision_tower.blocks.2.3.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
474 |
+
"vision_tower.blocks.2.3.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
475 |
+
"vision_tower.blocks.2.3.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
476 |
+
"vision_tower.blocks.2.3.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
477 |
+
"vision_tower.blocks.2.3.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
478 |
+
"vision_tower.blocks.2.3.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
479 |
+
"vision_tower.blocks.2.3.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
480 |
+
"vision_tower.blocks.2.3.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
481 |
+
"vision_tower.blocks.2.3.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
482 |
+
"vision_tower.blocks.2.3.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
483 |
+
"vision_tower.blocks.2.3.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
484 |
+
"vision_tower.blocks.2.3.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
485 |
+
"vision_tower.blocks.2.3.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
486 |
+
"vision_tower.blocks.2.3.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
487 |
+
"vision_tower.blocks.2.3.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
488 |
+
"vision_tower.blocks.2.3.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
489 |
+
"vision_tower.blocks.2.3.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
490 |
+
"vision_tower.blocks.2.3.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
491 |
+
"vision_tower.blocks.2.4.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
492 |
+
"vision_tower.blocks.2.4.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
493 |
+
"vision_tower.blocks.2.4.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
494 |
+
"vision_tower.blocks.2.4.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
495 |
+
"vision_tower.blocks.2.4.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
496 |
+
"vision_tower.blocks.2.4.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
497 |
+
"vision_tower.blocks.2.4.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
498 |
+
"vision_tower.blocks.2.4.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
499 |
+
"vision_tower.blocks.2.4.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
500 |
+
"vision_tower.blocks.2.4.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
501 |
+
"vision_tower.blocks.2.4.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
502 |
+
"vision_tower.blocks.2.4.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
503 |
+
"vision_tower.blocks.2.4.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
504 |
+
"vision_tower.blocks.2.4.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
505 |
+
"vision_tower.blocks.2.4.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
506 |
+
"vision_tower.blocks.2.4.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
507 |
+
"vision_tower.blocks.2.4.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
508 |
+
"vision_tower.blocks.2.4.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
509 |
+
"vision_tower.blocks.2.4.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
510 |
+
"vision_tower.blocks.2.4.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
511 |
+
"vision_tower.blocks.2.4.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
512 |
+
"vision_tower.blocks.2.4.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
513 |
+
"vision_tower.blocks.2.4.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
514 |
+
"vision_tower.blocks.2.4.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
515 |
+
"vision_tower.blocks.2.4.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
516 |
+
"vision_tower.blocks.2.4.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
517 |
+
"vision_tower.blocks.2.4.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
518 |
+
"vision_tower.blocks.2.4.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
519 |
+
"vision_tower.blocks.2.4.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
520 |
+
"vision_tower.blocks.2.4.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
521 |
+
"vision_tower.blocks.2.4.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
522 |
+
"vision_tower.blocks.2.4.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
523 |
+
"vision_tower.blocks.2.5.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
524 |
+
"vision_tower.blocks.2.5.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
525 |
+
"vision_tower.blocks.2.5.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
526 |
+
"vision_tower.blocks.2.5.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
527 |
+
"vision_tower.blocks.2.5.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
528 |
+
"vision_tower.blocks.2.5.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
529 |
+
"vision_tower.blocks.2.5.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
530 |
+
"vision_tower.blocks.2.5.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
531 |
+
"vision_tower.blocks.2.5.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
532 |
+
"vision_tower.blocks.2.5.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
533 |
+
"vision_tower.blocks.2.5.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
534 |
+
"vision_tower.blocks.2.5.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
535 |
+
"vision_tower.blocks.2.5.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
536 |
+
"vision_tower.blocks.2.5.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
537 |
+
"vision_tower.blocks.2.5.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
538 |
+
"vision_tower.blocks.2.5.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
539 |
+
"vision_tower.blocks.2.5.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
540 |
+
"vision_tower.blocks.2.5.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
541 |
+
"vision_tower.blocks.2.5.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
542 |
+
"vision_tower.blocks.2.5.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
543 |
+
"vision_tower.blocks.2.5.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
544 |
+
"vision_tower.blocks.2.5.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
545 |
+
"vision_tower.blocks.2.5.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
546 |
+
"vision_tower.blocks.2.5.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
547 |
+
"vision_tower.blocks.2.5.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
548 |
+
"vision_tower.blocks.2.5.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
549 |
+
"vision_tower.blocks.2.5.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
550 |
+
"vision_tower.blocks.2.5.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
551 |
+
"vision_tower.blocks.2.5.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
552 |
+
"vision_tower.blocks.2.5.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
553 |
+
"vision_tower.blocks.2.5.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
554 |
+
"vision_tower.blocks.2.5.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
555 |
+
"vision_tower.blocks.2.6.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
556 |
+
"vision_tower.blocks.2.6.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
557 |
+
"vision_tower.blocks.2.6.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
558 |
+
"vision_tower.blocks.2.6.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
559 |
+
"vision_tower.blocks.2.6.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
560 |
+
"vision_tower.blocks.2.6.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
561 |
+
"vision_tower.blocks.2.6.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
562 |
+
"vision_tower.blocks.2.6.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
563 |
+
"vision_tower.blocks.2.6.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
564 |
+
"vision_tower.blocks.2.6.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
565 |
+
"vision_tower.blocks.2.6.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
566 |
+
"vision_tower.blocks.2.6.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
567 |
+
"vision_tower.blocks.2.6.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
568 |
+
"vision_tower.blocks.2.6.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
569 |
+
"vision_tower.blocks.2.6.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
570 |
+
"vision_tower.blocks.2.6.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
571 |
+
"vision_tower.blocks.2.6.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
572 |
+
"vision_tower.blocks.2.6.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
573 |
+
"vision_tower.blocks.2.6.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
574 |
+
"vision_tower.blocks.2.6.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
575 |
+
"vision_tower.blocks.2.6.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
576 |
+
"vision_tower.blocks.2.6.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
577 |
+
"vision_tower.blocks.2.6.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
578 |
+
"vision_tower.blocks.2.6.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
579 |
+
"vision_tower.blocks.2.6.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
580 |
+
"vision_tower.blocks.2.6.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
581 |
+
"vision_tower.blocks.2.6.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
582 |
+
"vision_tower.blocks.2.6.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
583 |
+
"vision_tower.blocks.2.6.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
584 |
+
"vision_tower.blocks.2.6.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
585 |
+
"vision_tower.blocks.2.6.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
586 |
+
"vision_tower.blocks.2.6.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
587 |
+
"vision_tower.blocks.2.7.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
588 |
+
"vision_tower.blocks.2.7.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
589 |
+
"vision_tower.blocks.2.7.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
590 |
+
"vision_tower.blocks.2.7.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
591 |
+
"vision_tower.blocks.2.7.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
592 |
+
"vision_tower.blocks.2.7.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
593 |
+
"vision_tower.blocks.2.7.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
594 |
+
"vision_tower.blocks.2.7.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
595 |
+
"vision_tower.blocks.2.7.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
596 |
+
"vision_tower.blocks.2.7.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
597 |
+
"vision_tower.blocks.2.7.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
598 |
+
"vision_tower.blocks.2.7.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
599 |
+
"vision_tower.blocks.2.7.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
600 |
+
"vision_tower.blocks.2.7.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
601 |
+
"vision_tower.blocks.2.7.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
602 |
+
"vision_tower.blocks.2.7.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
603 |
+
"vision_tower.blocks.2.7.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
604 |
+
"vision_tower.blocks.2.7.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
605 |
+
"vision_tower.blocks.2.7.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
606 |
+
"vision_tower.blocks.2.7.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
607 |
+
"vision_tower.blocks.2.7.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
608 |
+
"vision_tower.blocks.2.7.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
609 |
+
"vision_tower.blocks.2.7.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
610 |
+
"vision_tower.blocks.2.7.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
611 |
+
"vision_tower.blocks.2.7.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
612 |
+
"vision_tower.blocks.2.7.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
613 |
+
"vision_tower.blocks.2.7.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
614 |
+
"vision_tower.blocks.2.7.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
615 |
+
"vision_tower.blocks.2.7.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
616 |
+
"vision_tower.blocks.2.7.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
617 |
+
"vision_tower.blocks.2.7.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
618 |
+
"vision_tower.blocks.2.7.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
619 |
+
"vision_tower.blocks.2.8.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
620 |
+
"vision_tower.blocks.2.8.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
621 |
+
"vision_tower.blocks.2.8.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
622 |
+
"vision_tower.blocks.2.8.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
623 |
+
"vision_tower.blocks.2.8.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
624 |
+
"vision_tower.blocks.2.8.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
625 |
+
"vision_tower.blocks.2.8.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
626 |
+
"vision_tower.blocks.2.8.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
627 |
+
"vision_tower.blocks.2.8.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
628 |
+
"vision_tower.blocks.2.8.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
629 |
+
"vision_tower.blocks.2.8.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
630 |
+
"vision_tower.blocks.2.8.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
631 |
+
"vision_tower.blocks.2.8.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
632 |
+
"vision_tower.blocks.2.8.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
633 |
+
"vision_tower.blocks.2.8.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
634 |
+
"vision_tower.blocks.2.8.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
635 |
+
"vision_tower.blocks.2.8.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
636 |
+
"vision_tower.blocks.2.8.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
637 |
+
"vision_tower.blocks.2.8.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
638 |
+
"vision_tower.blocks.2.8.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
639 |
+
"vision_tower.blocks.2.8.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
640 |
+
"vision_tower.blocks.2.8.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
641 |
+
"vision_tower.blocks.2.8.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
642 |
+
"vision_tower.blocks.2.8.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
643 |
+
"vision_tower.blocks.2.8.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
644 |
+
"vision_tower.blocks.2.8.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
645 |
+
"vision_tower.blocks.2.8.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
646 |
+
"vision_tower.blocks.2.8.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
647 |
+
"vision_tower.blocks.2.8.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
648 |
+
"vision_tower.blocks.2.8.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
649 |
+
"vision_tower.blocks.2.8.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
650 |
+
"vision_tower.blocks.2.8.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
651 |
+
"vision_tower.blocks.3.0.channel_block.channel_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
652 |
+
"vision_tower.blocks.3.0.channel_block.channel_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
653 |
+
"vision_tower.blocks.3.0.channel_block.channel_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
654 |
+
"vision_tower.blocks.3.0.channel_block.channel_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
655 |
+
"vision_tower.blocks.3.0.channel_block.channel_attn.norm.bias": "model-00003-of-00003.safetensors",
|
656 |
+
"vision_tower.blocks.3.0.channel_block.channel_attn.norm.weight": "model-00003-of-00003.safetensors",
|
657 |
+
"vision_tower.blocks.3.0.channel_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
658 |
+
"vision_tower.blocks.3.0.channel_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
659 |
+
"vision_tower.blocks.3.0.channel_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
660 |
+
"vision_tower.blocks.3.0.channel_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
661 |
+
"vision_tower.blocks.3.0.channel_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
662 |
+
"vision_tower.blocks.3.0.channel_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
663 |
+
"vision_tower.blocks.3.0.channel_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
664 |
+
"vision_tower.blocks.3.0.channel_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
665 |
+
"vision_tower.blocks.3.0.channel_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
666 |
+
"vision_tower.blocks.3.0.channel_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
667 |
+
"vision_tower.blocks.3.0.spatial_block.conv1.fn.dw.bias": "model-00003-of-00003.safetensors",
|
668 |
+
"vision_tower.blocks.3.0.spatial_block.conv1.fn.dw.weight": "model-00003-of-00003.safetensors",
|
669 |
+
"vision_tower.blocks.3.0.spatial_block.conv2.fn.dw.bias": "model-00003-of-00003.safetensors",
|
670 |
+
"vision_tower.blocks.3.0.spatial_block.conv2.fn.dw.weight": "model-00003-of-00003.safetensors",
|
671 |
+
"vision_tower.blocks.3.0.spatial_block.ffn.fn.net.fc1.bias": "model-00003-of-00003.safetensors",
|
672 |
+
"vision_tower.blocks.3.0.spatial_block.ffn.fn.net.fc1.weight": "model-00003-of-00003.safetensors",
|
673 |
+
"vision_tower.blocks.3.0.spatial_block.ffn.fn.net.fc2.bias": "model-00003-of-00003.safetensors",
|
674 |
+
"vision_tower.blocks.3.0.spatial_block.ffn.fn.net.fc2.weight": "model-00003-of-00003.safetensors",
|
675 |
+
"vision_tower.blocks.3.0.spatial_block.ffn.norm.bias": "model-00003-of-00003.safetensors",
|
676 |
+
"vision_tower.blocks.3.0.spatial_block.ffn.norm.weight": "model-00003-of-00003.safetensors",
|
677 |
+
"vision_tower.blocks.3.0.spatial_block.window_attn.fn.proj.bias": "model-00003-of-00003.safetensors",
|
678 |
+
"vision_tower.blocks.3.0.spatial_block.window_attn.fn.proj.weight": "model-00003-of-00003.safetensors",
|
679 |
+
"vision_tower.blocks.3.0.spatial_block.window_attn.fn.qkv.bias": "model-00003-of-00003.safetensors",
|
680 |
+
"vision_tower.blocks.3.0.spatial_block.window_attn.fn.qkv.weight": "model-00003-of-00003.safetensors",
|
681 |
+
"vision_tower.blocks.3.0.spatial_block.window_attn.norm.bias": "model-00003-of-00003.safetensors",
|
682 |
+
"vision_tower.blocks.3.0.spatial_block.window_attn.norm.weight": "model-00003-of-00003.safetensors",
|
683 |
+
"vision_tower.convs.0.norm.bias": "model-00003-of-00003.safetensors",
|
684 |
+
"vision_tower.convs.0.norm.weight": "model-00003-of-00003.safetensors",
|
685 |
+
"vision_tower.convs.0.proj.bias": "model-00003-of-00003.safetensors",
|
686 |
+
"vision_tower.convs.0.proj.weight": "model-00003-of-00003.safetensors",
|
687 |
+
"vision_tower.convs.1.norm.bias": "model-00003-of-00003.safetensors",
|
688 |
+
"vision_tower.convs.1.norm.weight": "model-00003-of-00003.safetensors",
|
689 |
+
"vision_tower.convs.1.proj.bias": "model-00003-of-00003.safetensors",
|
690 |
+
"vision_tower.convs.1.proj.weight": "model-00003-of-00003.safetensors",
|
691 |
+
"vision_tower.convs.2.norm.bias": "model-00003-of-00003.safetensors",
|
692 |
+
"vision_tower.convs.2.norm.weight": "model-00003-of-00003.safetensors",
|
693 |
+
"vision_tower.convs.2.proj.bias": "model-00003-of-00003.safetensors",
|
694 |
+
"vision_tower.convs.2.proj.weight": "model-00003-of-00003.safetensors",
|
695 |
+
"vision_tower.convs.3.norm.bias": "model-00003-of-00003.safetensors",
|
696 |
+
"vision_tower.convs.3.norm.weight": "model-00003-of-00003.safetensors",
|
697 |
+
"vision_tower.convs.3.proj.bias": "model-00003-of-00003.safetensors",
|
698 |
+
"vision_tower.convs.3.proj.weight": "model-00003-of-00003.safetensors",
|
699 |
+
"vision_tower.head.bias": "model-00003-of-00003.safetensors",
|
700 |
+
"vision_tower.head.weight": "model-00003-of-00003.safetensors",
|
701 |
+
"vision_tower.norms.bias": "model-00003-of-00003.safetensors",
|
702 |
+
"vision_tower.norms.weight": "model-00003-of-00003.safetensors",
|
703 |
+
"visual_temporal_embed.pos_idx_to_embed": "model-00003-of-00003.safetensors"
|
704 |
+
}
|
705 |
+
}
|
modeling_feynmodel.py
ADDED
@@ -0,0 +1,1528 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# modeling_fynmodel : Imed MAGROUNE / 2024 - 09
|
2 |
+
# original code from modeling_FeynModel
|
3 |
+
# add DaVit Vision Tower
|
4 |
+
#
|
5 |
+
# update generate forward function
|
6 |
+
#
|
7 |
+
# add lora adapters
|
8 |
+
#
|
9 |
+
# train on coco OD and vision reasoning
|
10 |
+
# train on ScenceQA
|
11 |
+
#
|
12 |
+
# todo add mamaba layer
|
13 |
+
#
|
14 |
+
# todo train on Arc-AGI
|
15 |
+
|
16 |
+
|
17 |
+
from transformers.modeling_utils import PreTrainedModel
|
18 |
+
from transformers.utils import (
|
19 |
+
ModelOutput,
|
20 |
+
add_start_docstrings,
|
21 |
+
add_start_docstrings_to_model_forward,
|
22 |
+
is_flash_attn_2_available,
|
23 |
+
logging,
|
24 |
+
replace_return_docstrings,
|
25 |
+
is_flash_attn_2_available,
|
26 |
+
is_flash_attn_greater_or_equal_2_10,
|
27 |
+
)
|
28 |
+
from transformers.activations import ACT2FN
|
29 |
+
from transformers.modeling_attn_mask_utils import (
|
30 |
+
_prepare_4d_attention_mask,
|
31 |
+
_prepare_4d_attention_mask_for_sdpa,
|
32 |
+
_prepare_4d_causal_attention_mask,
|
33 |
+
_prepare_4d_causal_attention_mask_for_sdpa,
|
34 |
+
)
|
35 |
+
from transformers.modeling_outputs import (
|
36 |
+
BaseModelOutput,
|
37 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
38 |
+
Seq2SeqLMOutput,
|
39 |
+
Seq2SeqModelOutput,
|
40 |
+
)
|
41 |
+
|
42 |
+
from transformers.cache_utils import Cache, HybridCache
|
43 |
+
from transformers.modeling_outputs import (
|
44 |
+
BaseModelOutputWithPast,
|
45 |
+
CausalLMOutputWithPast,
|
46 |
+
SequenceClassifierOutputWithPast,
|
47 |
+
TokenClassifierOutput,
|
48 |
+
)
|
49 |
+
|
50 |
+
from typing import List, Optional, Tuple, Union
|
51 |
+
|
52 |
+
from transformers.models.gemma2.modeling_gemma2 import Gemma2Model, Gemma2ForCausalLM,Gemma2DecoderLayer,Gemma2RMSNorm
|
53 |
+
from .configuration_feynmodel import FeynModelConfig,Florence2VisionConfig
|
54 |
+
|
55 |
+
from transformers import AutoProcessor, AutoTokenizer, AutoModelForCausalLM
|
56 |
+
import json
|
57 |
+
import math
|
58 |
+
import torch
|
59 |
+
from torch import nn
|
60 |
+
import torch.nn.functional as F
|
61 |
+
import logging
|
62 |
+
|
63 |
+
from transformers.utils import (
|
64 |
+
ModelOutput,
|
65 |
+
add_start_docstrings,
|
66 |
+
add_start_docstrings_to_model_forward,
|
67 |
+
is_flash_attn_2_available,
|
68 |
+
logging,
|
69 |
+
replace_return_docstrings,
|
70 |
+
is_flash_attn_2_available,
|
71 |
+
is_flash_attn_greater_or_equal_2_10,
|
72 |
+
)
|
73 |
+
|
74 |
+
from transformers.modeling_utils import PreTrainedModel
|
75 |
+
|
76 |
+
from collections import OrderedDict
|
77 |
+
from einops import rearrange
|
78 |
+
from timm.models.layers import DropPath, trunc_normal_
|
79 |
+
|
80 |
+
logger = logging.get_logger(__name__)
|
81 |
+
|
82 |
+
class MySequential(nn.Sequential):
|
83 |
+
def forward(self, *inputs):
|
84 |
+
for module in self._modules.values():
|
85 |
+
if type(inputs) == tuple:
|
86 |
+
inputs = module(*inputs)
|
87 |
+
else:
|
88 |
+
inputs = module(inputs)
|
89 |
+
return inputs
|
90 |
+
|
91 |
+
|
92 |
+
class PreNorm(nn.Module):
|
93 |
+
def __init__(self, norm, fn, drop_path=None):
|
94 |
+
super().__init__()
|
95 |
+
self.norm = norm
|
96 |
+
self.fn = fn
|
97 |
+
self.drop_path = drop_path
|
98 |
+
|
99 |
+
def forward(self, x, *args, **kwargs):
|
100 |
+
shortcut = x
|
101 |
+
if self.norm != None:
|
102 |
+
x, size = self.fn(self.norm(x), *args, **kwargs)
|
103 |
+
else:
|
104 |
+
x, size = self.fn(x, *args, **kwargs)
|
105 |
+
|
106 |
+
if self.drop_path:
|
107 |
+
x = self.drop_path(x)
|
108 |
+
|
109 |
+
x = shortcut + x
|
110 |
+
|
111 |
+
return x, size
|
112 |
+
|
113 |
+
|
114 |
+
class Mlp(nn.Module):
|
115 |
+
def __init__(
|
116 |
+
self,
|
117 |
+
in_features,
|
118 |
+
hidden_features=None,
|
119 |
+
out_features=None,
|
120 |
+
act_layer=nn.GELU,
|
121 |
+
):
|
122 |
+
super().__init__()
|
123 |
+
out_features = out_features or in_features
|
124 |
+
hidden_features = hidden_features or in_features
|
125 |
+
self.net = nn.Sequential(OrderedDict([
|
126 |
+
("fc1", nn.Linear(in_features, hidden_features)),
|
127 |
+
("act", act_layer()),
|
128 |
+
("fc2", nn.Linear(hidden_features, out_features))
|
129 |
+
]))
|
130 |
+
|
131 |
+
def forward(self, x, size):
|
132 |
+
return self.net(x), size
|
133 |
+
|
134 |
+
|
135 |
+
class DepthWiseConv2d(nn.Module):
|
136 |
+
def __init__(
|
137 |
+
self,
|
138 |
+
dim_in,
|
139 |
+
kernel_size,
|
140 |
+
padding,
|
141 |
+
stride,
|
142 |
+
bias=True,
|
143 |
+
):
|
144 |
+
super().__init__()
|
145 |
+
self.dw = nn.Conv2d(
|
146 |
+
dim_in, dim_in,
|
147 |
+
kernel_size=kernel_size,
|
148 |
+
padding=padding,
|
149 |
+
groups=dim_in,
|
150 |
+
stride=stride,
|
151 |
+
bias=bias
|
152 |
+
)
|
153 |
+
|
154 |
+
def forward(self, x, size):
|
155 |
+
B, N, C = x.shape
|
156 |
+
H, W = size
|
157 |
+
assert N == H * W
|
158 |
+
|
159 |
+
x = self.dw(x.transpose(1, 2).view(B, C, H, W))
|
160 |
+
size = (x.size(-2), x.size(-1))
|
161 |
+
x = x.flatten(2).transpose(1, 2)
|
162 |
+
return x, size
|
163 |
+
|
164 |
+
|
165 |
+
class ConvEmbed(nn.Module):
|
166 |
+
""" Image to Patch Embedding
|
167 |
+
"""
|
168 |
+
|
169 |
+
def __init__(
|
170 |
+
self,
|
171 |
+
patch_size=7,
|
172 |
+
in_chans=3,
|
173 |
+
embed_dim=64,
|
174 |
+
stride=4,
|
175 |
+
padding=2,
|
176 |
+
norm_layer=None,
|
177 |
+
pre_norm=True
|
178 |
+
):
|
179 |
+
super().__init__()
|
180 |
+
self.patch_size = patch_size
|
181 |
+
|
182 |
+
self.proj = nn.Conv2d(
|
183 |
+
in_chans, embed_dim,
|
184 |
+
kernel_size=patch_size,
|
185 |
+
stride=stride,
|
186 |
+
padding=padding
|
187 |
+
)
|
188 |
+
|
189 |
+
dim_norm = in_chans if pre_norm else embed_dim
|
190 |
+
self.norm = norm_layer(dim_norm) if norm_layer else None
|
191 |
+
|
192 |
+
self.pre_norm = pre_norm
|
193 |
+
|
194 |
+
def forward(self, x, size):
|
195 |
+
H, W = size
|
196 |
+
if len(x.size()) == 3:
|
197 |
+
if self.norm and self.pre_norm:
|
198 |
+
x = self.norm(x)
|
199 |
+
x = rearrange(
|
200 |
+
x, 'b (h w) c -> b c h w',
|
201 |
+
h=H, w=W
|
202 |
+
)
|
203 |
+
|
204 |
+
x = self.proj(x)
|
205 |
+
|
206 |
+
_, _, H, W = x.shape
|
207 |
+
x = rearrange(x, 'b c h w -> b (h w) c')
|
208 |
+
if self.norm and not self.pre_norm:
|
209 |
+
x = self.norm(x)
|
210 |
+
|
211 |
+
return x, (H, W)
|
212 |
+
|
213 |
+
|
214 |
+
class ChannelAttention(nn.Module):
|
215 |
+
|
216 |
+
def __init__(self, dim, groups=8, qkv_bias=True):
|
217 |
+
super().__init__()
|
218 |
+
|
219 |
+
self.groups = groups
|
220 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
221 |
+
self.proj = nn.Linear(dim, dim)
|
222 |
+
|
223 |
+
def forward(self, x, size):
|
224 |
+
B, N, C = x.shape
|
225 |
+
|
226 |
+
qkv = self.qkv(x).reshape(B, N, 3, self.groups, C // self.groups).permute(2, 0, 3, 1, 4)
|
227 |
+
q, k, v = qkv[0], qkv[1], qkv[2]
|
228 |
+
|
229 |
+
q = q * (float(N) ** -0.5)
|
230 |
+
attention = q.transpose(-1, -2) @ k
|
231 |
+
attention = attention.softmax(dim=-1)
|
232 |
+
x = (attention @ v.transpose(-1, -2)).transpose(-1, -2)
|
233 |
+
x = x.transpose(1, 2).reshape(B, N, C)
|
234 |
+
x = self.proj(x)
|
235 |
+
return x, size
|
236 |
+
|
237 |
+
|
238 |
+
class ChannelBlock(nn.Module):
|
239 |
+
|
240 |
+
def __init__(self, dim, groups, mlp_ratio=4., qkv_bias=True,
|
241 |
+
drop_path_rate=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm,
|
242 |
+
conv_at_attn=True, conv_at_ffn=True):
|
243 |
+
super().__init__()
|
244 |
+
|
245 |
+
drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
|
246 |
+
|
247 |
+
self.conv1 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_attn else None
|
248 |
+
self.channel_attn = PreNorm(
|
249 |
+
norm_layer(dim),
|
250 |
+
ChannelAttention(dim, groups=groups, qkv_bias=qkv_bias),
|
251 |
+
drop_path
|
252 |
+
)
|
253 |
+
self.conv2 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_ffn else None
|
254 |
+
self.ffn = PreNorm(
|
255 |
+
norm_layer(dim),
|
256 |
+
Mlp(in_features=dim, hidden_features=int(dim*mlp_ratio), act_layer=act_layer),
|
257 |
+
drop_path
|
258 |
+
)
|
259 |
+
|
260 |
+
def forward(self, x, size):
|
261 |
+
if self.conv1:
|
262 |
+
x, size = self.conv1(x, size)
|
263 |
+
x, size = self.channel_attn(x, size)
|
264 |
+
|
265 |
+
if self.conv2:
|
266 |
+
x, size = self.conv2(x, size)
|
267 |
+
x, size = self.ffn(x, size)
|
268 |
+
|
269 |
+
return x, size
|
270 |
+
|
271 |
+
|
272 |
+
def window_partition(x, window_size: int):
|
273 |
+
B, H, W, C = x.shape
|
274 |
+
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
|
275 |
+
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
|
276 |
+
return windows
|
277 |
+
|
278 |
+
|
279 |
+
def window_reverse(windows, batch_size: int, window_size: int, H: int, W: int):
|
280 |
+
B = batch_size
|
281 |
+
# this will cause onnx conversion failed for dynamic axis, because treated as constant
|
282 |
+
# int(windows.shape[0] / (H * W / window_size / window_size))
|
283 |
+
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
|
284 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
|
285 |
+
return x
|
286 |
+
|
287 |
+
|
288 |
+
class WindowAttention(nn.Module):
|
289 |
+
def __init__(self, dim, num_heads, window_size, qkv_bias=True):
|
290 |
+
|
291 |
+
super().__init__()
|
292 |
+
self.dim = dim
|
293 |
+
self.window_size = window_size
|
294 |
+
self.num_heads = num_heads
|
295 |
+
head_dim = dim // num_heads
|
296 |
+
self.scale = float(head_dim) ** -0.5
|
297 |
+
|
298 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
299 |
+
self.proj = nn.Linear(dim, dim)
|
300 |
+
|
301 |
+
self.softmax = nn.Softmax(dim=-1)
|
302 |
+
|
303 |
+
def forward(self, x, size):
|
304 |
+
|
305 |
+
H, W = size
|
306 |
+
B, L, C = x.shape
|
307 |
+
assert L == H * W, "input feature has wrong size"
|
308 |
+
|
309 |
+
x = x.view(B, H, W, C)
|
310 |
+
|
311 |
+
pad_l = pad_t = 0
|
312 |
+
pad_r = (self.window_size - W % self.window_size) % self.window_size
|
313 |
+
pad_b = (self.window_size - H % self.window_size) % self.window_size
|
314 |
+
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
|
315 |
+
_, Hp, Wp, _ = x.shape
|
316 |
+
|
317 |
+
x = window_partition(x, self.window_size)
|
318 |
+
x = x.view(-1, self.window_size * self.window_size, C)
|
319 |
+
|
320 |
+
# W-MSA/SW-MSA
|
321 |
+
# attn_windows = self.attn(x_windows)
|
322 |
+
|
323 |
+
B_, N, C = x.shape
|
324 |
+
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
325 |
+
q, k, v = qkv[0], qkv[1], qkv[2]
|
326 |
+
|
327 |
+
q = q * self.scale
|
328 |
+
attn = (q @ k.transpose(-2, -1))
|
329 |
+
attn = self.softmax(attn)
|
330 |
+
|
331 |
+
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
|
332 |
+
x = self.proj(x)
|
333 |
+
|
334 |
+
# merge windows
|
335 |
+
x = x.view(
|
336 |
+
-1, self.window_size, self.window_size, C
|
337 |
+
)
|
338 |
+
x = window_reverse(x, B, self.window_size, Hp, Wp)
|
339 |
+
|
340 |
+
if pad_r > 0 or pad_b > 0:
|
341 |
+
x = x[:, :H, :W, :].contiguous()
|
342 |
+
|
343 |
+
x = x.view(B, H * W, C)
|
344 |
+
|
345 |
+
return x, size
|
346 |
+
|
347 |
+
|
348 |
+
class SpatialBlock(nn.Module):
|
349 |
+
|
350 |
+
def __init__(self, dim, num_heads, window_size,
|
351 |
+
mlp_ratio=4., qkv_bias=True, drop_path_rate=0., act_layer=nn.GELU,
|
352 |
+
norm_layer=nn.LayerNorm, conv_at_attn=True, conv_at_ffn=True):
|
353 |
+
super().__init__()
|
354 |
+
|
355 |
+
drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
|
356 |
+
|
357 |
+
self.conv1 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_attn else None
|
358 |
+
self.window_attn = PreNorm(
|
359 |
+
norm_layer(dim),
|
360 |
+
WindowAttention(dim, num_heads, window_size, qkv_bias=qkv_bias),
|
361 |
+
drop_path
|
362 |
+
)
|
363 |
+
self.conv2 = PreNorm(None, DepthWiseConv2d(dim, 3, 1, 1)) if conv_at_ffn else None
|
364 |
+
self.ffn = PreNorm(
|
365 |
+
norm_layer(dim),
|
366 |
+
Mlp(in_features=dim, hidden_features=int(dim*mlp_ratio), act_layer=act_layer),
|
367 |
+
drop_path
|
368 |
+
)
|
369 |
+
|
370 |
+
def forward(self, x, size):
|
371 |
+
if self.conv1:
|
372 |
+
x, size = self.conv1(x, size)
|
373 |
+
x, size = self.window_attn(x, size)
|
374 |
+
|
375 |
+
if self.conv2:
|
376 |
+
x, size = self.conv2(x, size)
|
377 |
+
x, size = self.ffn(x, size)
|
378 |
+
return x, size
|
379 |
+
|
380 |
+
|
381 |
+
class DaViT(nn.Module):
|
382 |
+
""" DaViT: Dual-Attention Transformer
|
383 |
+
|
384 |
+
Args:
|
385 |
+
in_chans (int): Number of input image channels. Default: 3.
|
386 |
+
num_classes (int): Number of classes for classification head. Default: 1000.
|
387 |
+
patch_size (tuple(int)): Patch size of convolution in different stages. Default: (7, 2, 2, 2).
|
388 |
+
patch_stride (tuple(int)): Patch stride of convolution in different stages. Default: (4, 2, 2, 2).
|
389 |
+
patch_padding (tuple(int)): Patch padding of convolution in different stages. Default: (3, 0, 0, 0).
|
390 |
+
patch_prenorm (tuple(bool)): If True, perform norm before convlution layer. Default: (True, False, False, False).
|
391 |
+
embed_dims (tuple(int)): Patch embedding dimension in different stages. Default: (64, 128, 192, 256).
|
392 |
+
num_heads (tuple(int)): Number of spatial attention heads in different stages. Default: (4, 8, 12, 16).
|
393 |
+
num_groups (tuple(int)): Number of channel groups in different stages. Default: (4, 8, 12, 16).
|
394 |
+
window_size (int): Window size. Default: 7.
|
395 |
+
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
|
396 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True.
|
397 |
+
drop_path_rate (float): Stochastic depth rate. Default: 0.1.
|
398 |
+
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
|
399 |
+
enable_checkpoint (bool): If True, enable checkpointing. Default: False.
|
400 |
+
conv_at_attn (bool): If True, performe depthwise convolution before attention layer. Default: True.
|
401 |
+
conv_at_ffn (bool): If True, performe depthwise convolution before ffn layer. Default: True.
|
402 |
+
"""
|
403 |
+
|
404 |
+
def __init__(
|
405 |
+
self,
|
406 |
+
in_chans=3,
|
407 |
+
num_classes=1000,
|
408 |
+
depths=(1, 1, 3, 1),
|
409 |
+
patch_size=(7, 2, 2, 2),
|
410 |
+
patch_stride=(4, 2, 2, 2),
|
411 |
+
patch_padding=(3, 0, 0, 0),
|
412 |
+
patch_prenorm=(False, False, False, False),
|
413 |
+
embed_dims=(64, 128, 192, 256),
|
414 |
+
num_heads=(3, 6, 12, 24),
|
415 |
+
num_groups=(3, 6, 12, 24),
|
416 |
+
window_size=7,
|
417 |
+
mlp_ratio=4.,
|
418 |
+
qkv_bias=True,
|
419 |
+
drop_path_rate=0.1,
|
420 |
+
norm_layer=nn.LayerNorm,
|
421 |
+
enable_checkpoint=False,
|
422 |
+
conv_at_attn=True,
|
423 |
+
conv_at_ffn=True,
|
424 |
+
):
|
425 |
+
super().__init__()
|
426 |
+
|
427 |
+
self.num_classes = num_classes
|
428 |
+
self.embed_dims = embed_dims
|
429 |
+
self.num_heads = num_heads
|
430 |
+
self.num_groups = num_groups
|
431 |
+
self.num_stages = len(self.embed_dims)
|
432 |
+
self.enable_checkpoint = enable_checkpoint
|
433 |
+
assert self.num_stages == len(self.num_heads) == len(self.num_groups)
|
434 |
+
|
435 |
+
num_stages = len(embed_dims)
|
436 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths)*2)]
|
437 |
+
|
438 |
+
depth_offset = 0
|
439 |
+
convs = []
|
440 |
+
blocks = []
|
441 |
+
for i in range(num_stages):
|
442 |
+
conv_embed = ConvEmbed(
|
443 |
+
patch_size=patch_size[i],
|
444 |
+
stride=patch_stride[i],
|
445 |
+
padding=patch_padding[i],
|
446 |
+
in_chans=in_chans if i == 0 else self.embed_dims[i - 1],
|
447 |
+
embed_dim=self.embed_dims[i],
|
448 |
+
norm_layer=norm_layer,
|
449 |
+
pre_norm=patch_prenorm[i]
|
450 |
+
)
|
451 |
+
convs.append(conv_embed)
|
452 |
+
|
453 |
+
block = MySequential(
|
454 |
+
*[
|
455 |
+
MySequential(OrderedDict([
|
456 |
+
(
|
457 |
+
'spatial_block', SpatialBlock(
|
458 |
+
embed_dims[i],
|
459 |
+
num_heads[i],
|
460 |
+
window_size,
|
461 |
+
drop_path_rate=dpr[depth_offset+j*2],
|
462 |
+
qkv_bias=qkv_bias,
|
463 |
+
mlp_ratio=mlp_ratio,
|
464 |
+
conv_at_attn=conv_at_attn,
|
465 |
+
conv_at_ffn=conv_at_ffn,
|
466 |
+
)
|
467 |
+
),
|
468 |
+
(
|
469 |
+
'channel_block', ChannelBlock(
|
470 |
+
embed_dims[i],
|
471 |
+
num_groups[i],
|
472 |
+
drop_path_rate=dpr[depth_offset+j*2+1],
|
473 |
+
qkv_bias=qkv_bias,
|
474 |
+
mlp_ratio=mlp_ratio,
|
475 |
+
conv_at_attn=conv_at_attn,
|
476 |
+
conv_at_ffn=conv_at_ffn,
|
477 |
+
)
|
478 |
+
)
|
479 |
+
])) for j in range(depths[i])
|
480 |
+
]
|
481 |
+
)
|
482 |
+
blocks.append(block)
|
483 |
+
depth_offset += depths[i]*2
|
484 |
+
|
485 |
+
self.convs = nn.ModuleList(convs)
|
486 |
+
self.blocks = nn.ModuleList(blocks)
|
487 |
+
|
488 |
+
self.norms = norm_layer(self.embed_dims[-1])
|
489 |
+
self.avgpool = nn.AdaptiveAvgPool1d(1)
|
490 |
+
self.head = nn.Linear(self.embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity()
|
491 |
+
|
492 |
+
self.apply(self._init_weights)
|
493 |
+
|
494 |
+
@property
|
495 |
+
def dim_out(self):
|
496 |
+
return self.embed_dims[-1]
|
497 |
+
|
498 |
+
def _init_weights(self, m):
|
499 |
+
if isinstance(m, nn.Linear):
|
500 |
+
trunc_normal_(m.weight, std=0.02)
|
501 |
+
if m.bias is not None:
|
502 |
+
nn.init.constant_(m.bias, 0)
|
503 |
+
elif isinstance(m, nn.Conv2d):
|
504 |
+
nn.init.normal_(m.weight, std=0.02)
|
505 |
+
for name, _ in m.named_parameters():
|
506 |
+
if name in ['bias']:
|
507 |
+
nn.init.constant_(m.bias, 0)
|
508 |
+
elif isinstance(m, nn.LayerNorm):
|
509 |
+
nn.init.constant_(m.weight, 1.0)
|
510 |
+
nn.init.constant_(m.bias, 0)
|
511 |
+
elif isinstance(m, nn.BatchNorm2d):
|
512 |
+
nn.init.constant_(m.weight, 1.0)
|
513 |
+
nn.init.constant_(m.bias, 0)
|
514 |
+
|
515 |
+
def forward_features_unpool(self, x):
|
516 |
+
"""
|
517 |
+
forward until avg pooling
|
518 |
+
Args:
|
519 |
+
x (_type_): input image tensor
|
520 |
+
"""
|
521 |
+
input_size = (x.size(2), x.size(3))
|
522 |
+
for conv, block in zip(self.convs, self.blocks):
|
523 |
+
x, input_size = conv(x, input_size)
|
524 |
+
if self.enable_checkpoint:
|
525 |
+
x, input_size = checkpoint.checkpoint(block, x, input_size)
|
526 |
+
else:
|
527 |
+
x, input_size = block(x, input_size)
|
528 |
+
return x
|
529 |
+
|
530 |
+
def forward_features(self, x):
|
531 |
+
x = self.forward_features_unpool(x)
|
532 |
+
|
533 |
+
# (batch_size, num_tokens, token_dim)
|
534 |
+
x = self.avgpool(x.transpose(1, 2))
|
535 |
+
# (batch_size, 1, num_tokens)
|
536 |
+
x = torch.flatten(x, 1)
|
537 |
+
x = self.norms(x)
|
538 |
+
|
539 |
+
return x
|
540 |
+
|
541 |
+
def forward(self, x):
|
542 |
+
x = self.forward_features(x)
|
543 |
+
x = self.head(x)
|
544 |
+
return x
|
545 |
+
|
546 |
+
@classmethod
|
547 |
+
def from_config(cls, config):
|
548 |
+
return cls(
|
549 |
+
depths=config.depths,
|
550 |
+
embed_dims=config.dim_embed,
|
551 |
+
num_heads=config.num_heads,
|
552 |
+
num_groups=config.num_groups,
|
553 |
+
patch_size=config.patch_size,
|
554 |
+
patch_stride=config.patch_stride,
|
555 |
+
patch_padding=config.patch_padding,
|
556 |
+
patch_prenorm=config.patch_prenorm,
|
557 |
+
drop_path_rate=config.drop_path_rate,
|
558 |
+
window_size=config.window_size,
|
559 |
+
)
|
560 |
+
|
561 |
+
|
562 |
+
|
563 |
+
|
564 |
+
_CONFIG_FOR_DOC = "FeynModelConfig"
|
565 |
+
|
566 |
+
FEYNMODEL_START_DOCSTRING = r"""
|
567 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
568 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
569 |
+
etc.)
|
570 |
+
|
571 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
572 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
573 |
+
and behavior.
|
574 |
+
|
575 |
+
Parameters:
|
576 |
+
config ([`FeynModelConfig`]):
|
577 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
578 |
+
load the weights associated with the model, only the configuration. Check out the
|
579 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
580 |
+
"""
|
581 |
+
FEYNMODEL_INPUTS_DOCSTRING = r"""
|
582 |
+
Args:
|
583 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
584 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
585 |
+
it.
|
586 |
+
|
587 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
588 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
589 |
+
|
590 |
+
[What are input IDs?](../glossary#input-ids)
|
591 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
592 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
593 |
+
|
594 |
+
- 1 for tokens that are **not masked**,
|
595 |
+
- 0 for tokens that are **masked**.
|
596 |
+
|
597 |
+
[What are attention masks?](../glossary#attention-mask)
|
598 |
+
|
599 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
600 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
601 |
+
|
602 |
+
If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
|
603 |
+
`past_key_values`).
|
604 |
+
|
605 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
606 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
607 |
+
information on the default strategy.
|
608 |
+
|
609 |
+
- 1 indicates the head is **not masked**,
|
610 |
+
- 0 indicates the head is **masked**.
|
611 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
612 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
613 |
+
config.n_positions - 1]`.
|
614 |
+
|
615 |
+
[What are position IDs?](../glossary#position-ids)
|
616 |
+
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
617 |
+
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
618 |
+
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
619 |
+
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
620 |
+
|
621 |
+
Two formats are allowed:
|
622 |
+
- a [`~cache_utils.Cache`] instance;
|
623 |
+
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
624 |
+
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
625 |
+
cache format.
|
626 |
+
|
627 |
+
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
628 |
+
legacy cache format will be returned.
|
629 |
+
|
630 |
+
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
631 |
+
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
632 |
+
of shape `(batch_size, sequence_length)`.
|
633 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
634 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
635 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
636 |
+
model's internal embedding lookup matrix.
|
637 |
+
use_cache (`bool`, *optional*):
|
638 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
639 |
+
`past_key_values`).
|
640 |
+
output_attentions (`bool`, *optional*):
|
641 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
642 |
+
tensors for more detail.
|
643 |
+
output_hidden_states (`bool`, *optional*):
|
644 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
645 |
+
more detail.
|
646 |
+
return_dict (`bool`, *optional*):
|
647 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
648 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
649 |
+
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
|
650 |
+
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
|
651 |
+
the complete sequence length.
|
652 |
+
"""
|
653 |
+
|
654 |
+
# Copied from transformers.models.llama.modeling_llama._prepare_4d_causal_attention_mask_with_cache_position
|
655 |
+
def _prepare_4d_causal_attention_mask_with_cache_position(
|
656 |
+
attention_mask: torch.Tensor,
|
657 |
+
sequence_length: int,
|
658 |
+
target_length: int,
|
659 |
+
dtype: torch.dtype,
|
660 |
+
device: torch.device,
|
661 |
+
min_dtype: float,
|
662 |
+
cache_position: torch.Tensor,
|
663 |
+
batch_size: int,
|
664 |
+
):
|
665 |
+
|
666 |
+
#print(f" +++++++++ prepare 4K +++++++++++++++ rec {attention_mask.size()} sequence_length {sequence_length}")
|
667 |
+
if attention_mask is not None and attention_mask.dim() == 4:
|
668 |
+
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
669 |
+
#print("+++++++++++++++++ return it")
|
670 |
+
#causal_mask = attention_mask
|
671 |
+
# In this case we assume that the mask comes already in inverted form.
|
672 |
+
causal_mask = attention_mask[:, :, -sequence_length:, :]
|
673 |
+
#print(f"+++++++++++++++++ truncated causal_mask to last {sequence_length} elements, size: {causal_mask.size()}")
|
674 |
+
#print(f"+++++++++++++++++ return it causal_mask {causal_mask.size()} !!!!!!!!! attention_mask {attention_mask.size()}")
|
675 |
+
else:
|
676 |
+
#print("+++++++++++++++++++++ else +++++++++++++++++")
|
677 |
+
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
|
678 |
+
#print(f"++++++++++++++++ causal_mask {causal_mask.size()} ++++++++++++++++++ sequence_length = {sequence_length} ")
|
679 |
+
if sequence_length != 1:
|
680 |
+
causal_mask = torch.triu(causal_mask, diagonal=1)
|
681 |
+
#print(f"++++++++++++++++++ causal_mask = torch.triu ++++++++++ {causal_mask.size()} ")
|
682 |
+
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
683 |
+
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
684 |
+
#print(f"+++++++++++++++++++++ avant if attention_mask is not None:, causal_mask={causal_mask.size()}")
|
685 |
+
if attention_mask is not None:
|
686 |
+
#print(" +++++++++++++ attention_mask is None++++++++++++")
|
687 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
688 |
+
mask_length = attention_mask.shape[-1]
|
689 |
+
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
|
690 |
+
padding_mask = padding_mask == 0
|
691 |
+
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
692 |
+
padding_mask, min_dtype
|
693 |
+
)
|
694 |
+
#print(f"+++++++++++++++++++ 4K returning causal_mask {causal_mask.size()} +++++++++++++++++++")
|
695 |
+
|
696 |
+
return causal_mask
|
697 |
+
|
698 |
+
class LearnedAbsolutePositionEmbedding2D(nn.Module):
|
699 |
+
"""
|
700 |
+
This module learns positional embeddings up to a fixed maximum size.
|
701 |
+
"""
|
702 |
+
|
703 |
+
def __init__(self, embedding_dim=256, num_pos=50):
|
704 |
+
super().__init__()
|
705 |
+
self.row_embeddings = nn.Embedding(num_pos, embedding_dim // 2)
|
706 |
+
self.column_embeddings = nn.Embedding(num_pos, embedding_dim - (embedding_dim // 2))
|
707 |
+
|
708 |
+
def forward(self, pixel_values):
|
709 |
+
"""
|
710 |
+
pixel_values: (batch_size, height, width, num_channels)
|
711 |
+
returns: (batch_size, height, width, embedding_dim * 2)
|
712 |
+
"""
|
713 |
+
if len(pixel_values.shape) != 4:
|
714 |
+
raise ValueError('pixel_values must be a 4D tensor')
|
715 |
+
height, width = pixel_values.shape[1:3]
|
716 |
+
width_values = torch.arange(width, device=pixel_values.device)
|
717 |
+
height_values = torch.arange(height, device=pixel_values.device)
|
718 |
+
x_emb = self.column_embeddings(width_values)
|
719 |
+
y_emb = self.row_embeddings(height_values)
|
720 |
+
# (height, width, embedding_dim * 2)
|
721 |
+
pos = torch.cat([x_emb.unsqueeze(0).repeat(height, 1, 1), y_emb.unsqueeze(1).repeat(1, width, 1)], dim=-1)
|
722 |
+
# (embedding_dim * 2, height, width)
|
723 |
+
pos = pos.permute(2, 0, 1)
|
724 |
+
pos = pos.unsqueeze(0)
|
725 |
+
# (batch_size, embedding_dim * 2, height, width)
|
726 |
+
pos = pos.repeat(pixel_values.shape[0], 1, 1, 1)
|
727 |
+
# (batch_size, height, width, embedding_dim * 2)
|
728 |
+
pos = pos.permute(0, 2, 3, 1)
|
729 |
+
return pos
|
730 |
+
|
731 |
+
class PositionalEmbeddingCosine1D(nn.Module):
|
732 |
+
"""
|
733 |
+
This class implements a very simple positional encoding. It follows closely
|
734 |
+
the encoder from the link below:
|
735 |
+
https://pytorch.org/tutorials/beginner/translation_transformer.html
|
736 |
+
Args:
|
737 |
+
embed_dim: The dimension of the embeddings.
|
738 |
+
dropout_prob: The dropout probability.
|
739 |
+
max_seq_len: The maximum length to precompute the positional encodings.
|
740 |
+
"""
|
741 |
+
def __init__(
|
742 |
+
self,
|
743 |
+
embed_dim: int = 512,
|
744 |
+
max_seq_len: int = 1024) -> None:
|
745 |
+
super(PositionalEmbeddingCosine1D, self).__init__()
|
746 |
+
self.embed_dim = embed_dim
|
747 |
+
self.max_seq_len = max_seq_len
|
748 |
+
# Generate the sinusoidal arrays.
|
749 |
+
factor = math.log(10000)
|
750 |
+
denominator = torch.exp(
|
751 |
+
-factor * torch.arange(0, self.embed_dim, 2) / self.embed_dim)
|
752 |
+
# Matrix where rows correspond to a positional embedding as a function
|
753 |
+
# of the position index (i.e., the row index).
|
754 |
+
frequencies = \
|
755 |
+
torch.arange(0, self.max_seq_len) \
|
756 |
+
.reshape(self.max_seq_len, 1) * denominator
|
757 |
+
pos_idx_to_embed = torch.zeros((self.max_seq_len, self.embed_dim))
|
758 |
+
# Populate uneven entries.
|
759 |
+
pos_idx_to_embed[:, 0::2] = torch.sin(frequencies)
|
760 |
+
pos_idx_to_embed[:, 1::2] = torch.cos(frequencies)
|
761 |
+
# Save the positional embeddings in a constant buffer.
|
762 |
+
self.register_buffer("pos_idx_to_embed", pos_idx_to_embed)
|
763 |
+
|
764 |
+
def forward(self, seq_embeds: torch.Tensor) -> torch.Tensor:
|
765 |
+
"""
|
766 |
+
Args:
|
767 |
+
seq_embeds: The sequence embeddings in order. Allowed size:
|
768 |
+
1. [T, D], where T is the length of the sequence, and D is the
|
769 |
+
frame embedding dimension.
|
770 |
+
2. [B, T, D], where B is the batch size and T and D are the
|
771 |
+
same as above.
|
772 |
+
Returns a tensor of with the same dimensions as the input: i.e.,
|
773 |
+
[1, T, D] or [T, D].
|
774 |
+
"""
|
775 |
+
shape_len = len(seq_embeds.shape)
|
776 |
+
assert 2 <= shape_len <= 3
|
777 |
+
len_seq = seq_embeds.size(-2)
|
778 |
+
assert len_seq <= self.max_seq_len
|
779 |
+
pos_embeds = self.pos_idx_to_embed[0:seq_embeds.size(-2), :]
|
780 |
+
# Adapt pre-computed positional embeddings to the input.
|
781 |
+
if shape_len == 3:
|
782 |
+
pos_embeds = pos_embeds.view(
|
783 |
+
(1, pos_embeds.size(0), pos_embeds.size(1)))
|
784 |
+
return pos_embeds
|
785 |
+
|
786 |
+
|
787 |
+
class LearnedAbsolutePositionEmbedding1D(nn.Module):
|
788 |
+
"""
|
789 |
+
Learnable absolute positional embeddings for 1D sequences.
|
790 |
+
Args:
|
791 |
+
embed_dim: The dimension of the embeddings.
|
792 |
+
max_seq_len: The maximum length to precompute the positional encodings.
|
793 |
+
"""
|
794 |
+
def __init__(
|
795 |
+
self,
|
796 |
+
embedding_dim: int = 512,
|
797 |
+
num_pos: int = 1024) -> None:
|
798 |
+
super(LearnedAbsolutePositionEmbedding1D, self).__init__()
|
799 |
+
self.embeddings = nn.Embedding(num_pos, embedding_dim)
|
800 |
+
self.num_pos = num_pos
|
801 |
+
|
802 |
+
def forward(self, seq_embeds: torch.Tensor) -> torch.Tensor:
|
803 |
+
"""
|
804 |
+
Args:
|
805 |
+
seq_embeds: The sequence embeddings in order. Allowed size:
|
806 |
+
1. [T, D], where T is the length of the sequence, and D is the
|
807 |
+
frame embedding dimension.
|
808 |
+
2. [B, T, D], where B is the batch size and T and D are the
|
809 |
+
same as above.
|
810 |
+
Returns a tensor of with the same dimensions as the input: i.e.,
|
811 |
+
[1, T, D] or [T, D].
|
812 |
+
"""
|
813 |
+
shape_len = len(seq_embeds.shape)
|
814 |
+
assert 2 <= shape_len <= 3
|
815 |
+
len_seq = seq_embeds.size(-2)
|
816 |
+
assert len_seq <= self.num_pos
|
817 |
+
# [T, D]
|
818 |
+
pos_embeds = self.embeddings(torch.arange(len_seq).to(seq_embeds.device))
|
819 |
+
# Adapt pre-computed positional embeddings to the input.
|
820 |
+
if shape_len == 3:
|
821 |
+
pos_embeds = pos_embeds.view(
|
822 |
+
(1, pos_embeds.size(0), pos_embeds.size(1)))
|
823 |
+
return pos_embeds
|
824 |
+
|
825 |
+
def create_git_attention_mask(
|
826 |
+
tgt: torch.Tensor,
|
827 |
+
memory: torch.Tensor,
|
828 |
+
max_length: int
|
829 |
+
) -> torch.Tensor:
|
830 |
+
# Obtain the dimensions of the target text and memory
|
831 |
+
batch_size = tgt.size(0)
|
832 |
+
num_tgt = tgt.shape[1]
|
833 |
+
num_memory = memory.shape[1]
|
834 |
+
total_length = num_memory + num_tgt
|
835 |
+
|
836 |
+
# Create the top left part of the attention matrix
|
837 |
+
top_left = torch.zeros((num_memory, num_memory)) # Attention enabled in this region
|
838 |
+
top_right = torch.full((num_memory, num_tgt), float(-3.4028e+38)) # Attention disabled here
|
839 |
+
|
840 |
+
# Bottom left part of the attention matrix
|
841 |
+
bottom_left = torch.zeros((num_tgt, num_memory)) # Attention enabled here
|
842 |
+
|
843 |
+
# Create a lower triangular matrix for the bottom right part
|
844 |
+
bottom_right = torch.tril(torch.ones(num_tgt, num_tgt))
|
845 |
+
|
846 |
+
# Transform 1s to 0 to enable attention, and 0s to -inf to block attention
|
847 |
+
bottom_right = bottom_right.masked_fill(bottom_right == 0, float(-3.4028e+38))
|
848 |
+
bottom_right = bottom_right.masked_fill(bottom_right == 1, float(0))
|
849 |
+
|
850 |
+
# Concatenate matrices to form the full mask
|
851 |
+
left = torch.cat((top_left, bottom_left), dim=0)
|
852 |
+
right = torch.cat((top_right, bottom_right), dim=0)
|
853 |
+
|
854 |
+
# Combine left and right parts
|
855 |
+
full_attention_mask = torch.cat((left, right), dim=1)
|
856 |
+
|
857 |
+
# Add padding to reach max_length
|
858 |
+
padding = torch.full((total_length, max_length - total_length), float(-3.4028e+38))
|
859 |
+
full_attention_mask = torch.cat((full_attention_mask, padding), dim=1)
|
860 |
+
|
861 |
+
# Add an axis for multi-heads and batch_size
|
862 |
+
full_attention_mask = full_attention_mask[None, None, :, :]
|
863 |
+
|
864 |
+
# Expand the mask to have shape (batch_size, 1, seq_length, max_length)
|
865 |
+
full_attention_mask = full_attention_mask.expand(batch_size, 1, full_attention_mask.size(-2), full_attention_mask.size(-1))
|
866 |
+
|
867 |
+
return full_attention_mask
|
868 |
+
|
869 |
+
def get_position_ids_from_binary_attention_mask(mask):
|
870 |
+
"""
|
871 |
+
Extract position IDs from a binary attention mask.
|
872 |
+
|
873 |
+
Args:
|
874 |
+
mask (torch.Tensor): The attention mask tensor of shape (1, 1, seq_len, seq_len),
|
875 |
+
where 1 indicates allowed attention and 0 indicates blocked attention.
|
876 |
+
|
877 |
+
Returns:
|
878 |
+
list: A list of lists where each sublist contains the allowed position IDs for each query position.
|
879 |
+
"""
|
880 |
+
# Assuming the mask is of shape (1, 1, seq_len, seq_len)
|
881 |
+
_, _, seq_len, _ = mask.shape
|
882 |
+
|
883 |
+
# Create a tensor with position IDs from 0 to seq_len - 1
|
884 |
+
position_ids = torch.arange(seq_len, dtype=torch.long, device=mask.device)
|
885 |
+
|
886 |
+
# Add a batch dimension
|
887 |
+
position_ids = position_ids.unsqueeze(0)
|
888 |
+
|
889 |
+
return position_ids
|
890 |
+
|
891 |
+
def ensure_tensor(variable):
|
892 |
+
# Check if the variable is a torch.Tensor
|
893 |
+
if isinstance(variable, torch.Tensor):
|
894 |
+
# print("Variable is already a tensor.")
|
895 |
+
return variable
|
896 |
+
else:
|
897 |
+
#print("Variable is not a tensor, converting...")
|
898 |
+
try:
|
899 |
+
# Convert the variable to a tensor
|
900 |
+
tensor = torch.tensor(variable)
|
901 |
+
#print("Conversion successful.")
|
902 |
+
return tensor
|
903 |
+
except Exception as e:
|
904 |
+
print(f"Error converting to tensor: {e}")
|
905 |
+
raise
|
906 |
+
|
907 |
+
@add_start_docstrings(
|
908 |
+
"The bare Model outputting raw hidden-states without any specific head on top.",
|
909 |
+
FEYNMODEL_START_DOCSTRING,
|
910 |
+
)
|
911 |
+
class FeynModel(Gemma2Model):
|
912 |
+
"""
|
913 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers.
|
914 |
+
Each layer is a [`FeynModelDecoderLayer`] + ['LoraLayer'] for *proj* moduls
|
915 |
+
NB : LoraLayers will be added and activatd on proj modules onpy if pixel_values is not None
|
916 |
+
|
917 |
+
Args:
|
918 |
+
config: FeynModelConfig
|
919 |
+
"""
|
920 |
+
|
921 |
+
def __init__(self, config: FeynModelConfig):
|
922 |
+
super().__init__(config)
|
923 |
+
# Initialize weights and apply final processing
|
924 |
+
self.mode='llm'
|
925 |
+
'''
|
926 |
+
self.image_patch_tokens = int(
|
927 |
+
(config.vision_config.image_size / config.vision_config.patch_size) ** 2 + 1
|
928 |
+
)
|
929 |
+
|
930 |
+
if config.num_image_with_embedding is not None:
|
931 |
+
self.image_patch_tokens *= config.num_image_with_embedding
|
932 |
+
'''
|
933 |
+
self.image_patch_tokens = 577
|
934 |
+
self.post_init()
|
935 |
+
|
936 |
+
def get_input_embeddings(self):
|
937 |
+
return self.embed_tokens
|
938 |
+
|
939 |
+
def set_input_embeddings(self, value):
|
940 |
+
self.embed_tokens = value
|
941 |
+
|
942 |
+
|
943 |
+
|
944 |
+
|
945 |
+
@add_start_docstrings_to_model_forward(FEYNMODEL_INPUTS_DOCSTRING)
|
946 |
+
def forward(
|
947 |
+
self,
|
948 |
+
input_ids: torch.LongTensor = None,
|
949 |
+
attention_mask: Optional[torch.Tensor] = None,
|
950 |
+
position_ids: Optional[torch.LongTensor] = None,
|
951 |
+
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
952 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
953 |
+
use_cache: Optional[bool] = None,
|
954 |
+
output_attentions: Optional[bool] = None,
|
955 |
+
output_hidden_states: Optional[bool] = None,
|
956 |
+
return_dict: Optional[bool] = None,
|
957 |
+
cache_position: Optional[torch.LongTensor] = None,
|
958 |
+
causal_attention_mask: Optional[torch.Tensor] = None,
|
959 |
+
**kwargs,
|
960 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
961 |
+
|
962 |
+
# print(f" self.mode = {self.mode}")
|
963 |
+
# Ensure cache_position is initialized if not provided
|
964 |
+
|
965 |
+
|
966 |
+
if cache_position is None:
|
967 |
+
batch_size = input_ids.size(0) if input_ids is not None else inputs_embeds.size(0)
|
968 |
+
cache_position = torch.zeros((batch_size,), dtype=torch.long, device=input_ids.device if input_ids is not None else inputs_embeds.device)
|
969 |
+
|
970 |
+
|
971 |
+
|
972 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
973 |
+
output_hidden_states = (
|
974 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
975 |
+
)
|
976 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
977 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
978 |
+
|
979 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
980 |
+
raise ValueError(
|
981 |
+
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
|
982 |
+
)
|
983 |
+
|
984 |
+
if self.gradient_checkpointing and self.training and use_cache:
|
985 |
+
logger.warning_once(
|
986 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
987 |
+
)
|
988 |
+
use_cache = False
|
989 |
+
|
990 |
+
if inputs_embeds is None:
|
991 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
992 |
+
causal_mask = self._update_causal_mask(
|
993 |
+
attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
|
994 |
+
)
|
995 |
+
else:
|
996 |
+
causal_mask = ensure_tensor(causal_attention_mask)
|
997 |
+
position_ids = get_position_ids_from_binary_attention_mask(attention_mask)
|
998 |
+
|
999 |
+
#print(f" causal_mask = {causal_mask} ")
|
1000 |
+
|
1001 |
+
if cache_position is None:
|
1002 |
+
cache_position = torch.arange(0, inputs_embeds.shape[1], device=inputs_embeds.device)
|
1003 |
+
|
1004 |
+
if position_ids is None :
|
1005 |
+
position_ids = cache_position.unsqueeze(0)
|
1006 |
+
|
1007 |
+
|
1008 |
+
|
1009 |
+
# Convert position_ids to a tensor if not already
|
1010 |
+
if not isinstance(position_ids, torch.Tensor):
|
1011 |
+
|
1012 |
+
position_ids = torch.tensor(position_ids, dtype=torch.long, device=inputs_embeds.device)
|
1013 |
+
|
1014 |
+
|
1015 |
+
# embed positions
|
1016 |
+
hidden_states = inputs_embeds
|
1017 |
+
|
1018 |
+
# normalized
|
1019 |
+
# FeynModel downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
|
1020 |
+
# See https://github.com/huggingface/transformers/pull/29402
|
1021 |
+
normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
|
1022 |
+
hidden_states = hidden_states * normalizer
|
1023 |
+
|
1024 |
+
all_hidden_states = () if output_hidden_states else None
|
1025 |
+
all_self_attns = () if output_attentions else None
|
1026 |
+
|
1027 |
+
for decoder_layer in self.layers:
|
1028 |
+
if output_hidden_states:
|
1029 |
+
all_hidden_states += (hidden_states,)
|
1030 |
+
|
1031 |
+
if self.gradient_checkpointing and self.training:
|
1032 |
+
layer_outputs = self._gradient_checkpointing_func(
|
1033 |
+
decoder_layer.__call__,
|
1034 |
+
hidden_states,
|
1035 |
+
causal_mask,
|
1036 |
+
position_ids,
|
1037 |
+
past_key_values,
|
1038 |
+
output_attentions,
|
1039 |
+
use_cache,
|
1040 |
+
cache_position,
|
1041 |
+
)
|
1042 |
+
else:
|
1043 |
+
layer_outputs = decoder_layer(
|
1044 |
+
hidden_states,
|
1045 |
+
attention_mask=causal_mask,
|
1046 |
+
position_ids=position_ids,
|
1047 |
+
past_key_value=past_key_values,
|
1048 |
+
output_attentions=output_attentions,
|
1049 |
+
use_cache=use_cache,
|
1050 |
+
cache_position=cache_position,
|
1051 |
+
)
|
1052 |
+
|
1053 |
+
hidden_states = layer_outputs[0]
|
1054 |
+
|
1055 |
+
if output_attentions:
|
1056 |
+
all_self_attns += (layer_outputs[1],)
|
1057 |
+
|
1058 |
+
hidden_states = self.norm(hidden_states)
|
1059 |
+
|
1060 |
+
# add hidden states from the last decoder layer
|
1061 |
+
if output_hidden_states:
|
1062 |
+
all_hidden_states += (hidden_states,)
|
1063 |
+
|
1064 |
+
next_cache = past_key_values if use_cache else None
|
1065 |
+
|
1066 |
+
if not return_dict:
|
1067 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
1068 |
+
return BaseModelOutputWithPast(
|
1069 |
+
last_hidden_state=hidden_states,
|
1070 |
+
past_key_values=next_cache,
|
1071 |
+
hidden_states=all_hidden_states,
|
1072 |
+
attentions=all_self_attns,
|
1073 |
+
)
|
1074 |
+
|
1075 |
+
|
1076 |
+
|
1077 |
+
def _update_causal_mask(
|
1078 |
+
self,
|
1079 |
+
attention_mask: torch.Tensor,
|
1080 |
+
input_tensor: torch.Tensor,
|
1081 |
+
cache_position: torch.Tensor,
|
1082 |
+
past_key_values: Cache,
|
1083 |
+
output_attentions: bool,
|
1084 |
+
):
|
1085 |
+
|
1086 |
+
# print(f" _start _____ _update_causal_mask attention_mask {attention_mask.size()} {attention_mask} ")
|
1087 |
+
# Flash Attention currently doesn't support static cache but FeynModel work only with static cache.
|
1088 |
+
# So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape
|
1089 |
+
# to cut out keys/values trailing 0 used in static cache. This workaround should be compile compatible
|
1090 |
+
# as it doesn't cause dynamic control issues.
|
1091 |
+
if self.config._attn_implementation == "flash_attention_2":
|
1092 |
+
return attention_mask
|
1093 |
+
|
1094 |
+
dtype, device = input_tensor.dtype, input_tensor.device
|
1095 |
+
min_dtype = torch.finfo(dtype).min
|
1096 |
+
sequence_length = input_tensor.shape[1]
|
1097 |
+
if isinstance(past_key_values, HybridCache):
|
1098 |
+
target_length = past_key_values.get_max_length()
|
1099 |
+
else:
|
1100 |
+
target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]
|
1101 |
+
|
1102 |
+
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
1103 |
+
causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
|
1104 |
+
attention_mask,
|
1105 |
+
sequence_length=sequence_length,
|
1106 |
+
target_length=target_length,
|
1107 |
+
dtype=dtype,
|
1108 |
+
device=device,
|
1109 |
+
min_dtype=min_dtype,
|
1110 |
+
cache_position=cache_position,
|
1111 |
+
batch_size=input_tensor.shape[0],
|
1112 |
+
)
|
1113 |
+
#print(f" _end ______ _update_causal_mask causal_mask {causal_mask.size()} {causal_mask} ")
|
1114 |
+
return causal_mask
|
1115 |
+
|
1116 |
+
|
1117 |
+
|
1118 |
+
class FeynModelForCausalLM(Gemma2ForCausalLM):
|
1119 |
+
_tied_weights_keys = ["lm_head.weight"]
|
1120 |
+
config_class = FeynModelConfig
|
1121 |
+
def __init__(self, config):
|
1122 |
+
super().__init__(config)
|
1123 |
+
config.vision_config=Florence2VisionConfig.from_dict(config.vision_config)
|
1124 |
+
self.model = FeynModel(config)
|
1125 |
+
|
1126 |
+
# assert config.vision_config.model_type== 'davit', 'only DaViT is supported for now'
|
1127 |
+
self.vision_tower = DaViT.from_config(config=config.vision_config)
|
1128 |
+
self._build_image_projection_layers(config)
|
1129 |
+
|
1130 |
+
self.__causal_attention_mask = None
|
1131 |
+
|
1132 |
+
# Initialize weights and apply final processing
|
1133 |
+
self.post_init()
|
1134 |
+
|
1135 |
+
################ Vision Tower ########################
|
1136 |
+
def _build_image_projection_layers(self, config):
|
1137 |
+
image_dim_out = config.vision_config.dim_embed[-1]
|
1138 |
+
dim_projection = config.vision_config.projection_dim
|
1139 |
+
self.image_projection = nn.Parameter(
|
1140 |
+
torch.empty(image_dim_out, dim_projection)
|
1141 |
+
)
|
1142 |
+
self.image_proj_norm = nn.LayerNorm(dim_projection)
|
1143 |
+
image_pos_embed_config = config.vision_config.image_pos_embed
|
1144 |
+
if image_pos_embed_config['type'] == 'learned_abs_2d':
|
1145 |
+
self.image_pos_embed = LearnedAbsolutePositionEmbedding2D(
|
1146 |
+
embedding_dim=image_dim_out,
|
1147 |
+
num_pos=image_pos_embed_config['max_pos_embeddings']
|
1148 |
+
)
|
1149 |
+
else:
|
1150 |
+
raise NotImplementedError('Not implemented yet')
|
1151 |
+
|
1152 |
+
self.image_feature_source = config.vision_config.image_feature_source
|
1153 |
+
|
1154 |
+
# temporal embedding
|
1155 |
+
visual_temporal_embedding_config = config.vision_config.visual_temporal_embedding
|
1156 |
+
if visual_temporal_embedding_config['type'] == 'COSINE':
|
1157 |
+
self.visual_temporal_embed = PositionalEmbeddingCosine1D(
|
1158 |
+
embed_dim=image_dim_out,
|
1159 |
+
max_seq_len=visual_temporal_embedding_config['max_temporal_embeddings']
|
1160 |
+
)
|
1161 |
+
else:
|
1162 |
+
raise NotImplementedError('Not implemented yet')
|
1163 |
+
|
1164 |
+
|
1165 |
+
|
1166 |
+
def _merge_input_ids_with_image_features(self, image_features, inputs_embeds):
|
1167 |
+
batch_size, image_token_length = image_features.size()[:-1]
|
1168 |
+
device = image_features.device
|
1169 |
+
image_attention_mask = torch.ones(batch_size, image_token_length, device=device)
|
1170 |
+
|
1171 |
+
if inputs_embeds is None:
|
1172 |
+
return image_features, image_attention_mask
|
1173 |
+
|
1174 |
+
task_prefix_embeds = inputs_embeds
|
1175 |
+
task_prefix_attention_mask = torch.ones(batch_size, task_prefix_embeds.size(1), device=device)
|
1176 |
+
|
1177 |
+
# Assurer que les masques d'attention sont de deux dimensions
|
1178 |
+
if len(task_prefix_attention_mask.shape) == 3:
|
1179 |
+
task_prefix_attention_mask = task_prefix_attention_mask.squeeze(1)
|
1180 |
+
|
1181 |
+
# Vérifier la dimension de batch et ajuster si nécessaire
|
1182 |
+
if image_features.size(0) != task_prefix_embeds.size(0):
|
1183 |
+
raise ValueError("Batch sizes of image_features and task_prefix_embeds do not match")
|
1184 |
+
|
1185 |
+
# Ajouter une dimension fictive si les dimensions ne sont pas alignées
|
1186 |
+
if image_features.dim() < task_prefix_embeds.dim():
|
1187 |
+
image_features = image_features.unsqueeze(-1)
|
1188 |
+
elif task_prefix_embeds.dim() < image_features.dim():
|
1189 |
+
task_prefix_embeds = task_prefix_embeds.unsqueeze(-1)
|
1190 |
+
|
1191 |
+
# Assurer que toutes les dimensions, sauf dim=1, sont identiques
|
1192 |
+
if image_features.size(2) != task_prefix_embeds.size(2):
|
1193 |
+
# Ajuster ou signaler une erreur si les dimensions internes ne sont pas compatibles
|
1194 |
+
raise ValueError("Internal dimensions of image_features and task_prefix_embeds do not match")
|
1195 |
+
|
1196 |
+
inputs_embeds = torch.cat([image_features, task_prefix_embeds], dim=1)
|
1197 |
+
attention_mask = torch.cat([image_attention_mask, task_prefix_attention_mask], dim=1)
|
1198 |
+
|
1199 |
+
return inputs_embeds, attention_mask
|
1200 |
+
|
1201 |
+
def _encode_image(self, pixel_values):
|
1202 |
+
if len(pixel_values.shape) == 4:
|
1203 |
+
batch_size, C, H, W = pixel_values.shape
|
1204 |
+
T = 1
|
1205 |
+
x = self.vision_tower.forward_features_unpool(pixel_values)
|
1206 |
+
else:
|
1207 |
+
# Ajoute une dimension de batch au début si 'pixel_values' n'a que 3 dimensions (C, H, W)
|
1208 |
+
pixel_values = pixel_values.unsqueeze(0) # Ajoute une dimension de batch
|
1209 |
+
batch_size, C, H, W = pixel_values.shape
|
1210 |
+
T = 1
|
1211 |
+
x = self.vision_tower.forward_features_unpool(pixel_values)
|
1212 |
+
|
1213 |
+
if self.image_pos_embed is not None:
|
1214 |
+
x = x.view(batch_size * T, -1, x.shape[-1])
|
1215 |
+
num_tokens = x.shape[-2]
|
1216 |
+
h, w = int(num_tokens ** 0.5), int(num_tokens ** 0.5)
|
1217 |
+
assert h * w == num_tokens, 'only support square feature maps for now'
|
1218 |
+
x = x.view(batch_size * T, h, w, x.shape[-1])
|
1219 |
+
pos_embed = self.image_pos_embed(x)
|
1220 |
+
x = x + pos_embed
|
1221 |
+
x = x.view(batch_size, T * h*w, x.shape[-1])
|
1222 |
+
|
1223 |
+
if self.visual_temporal_embed is not None:
|
1224 |
+
visual_temporal_embed = self.visual_temporal_embed(x.view(batch_size, T, -1, x.shape[-1])[:, :, 0])
|
1225 |
+
x = x.view(batch_size, T, -1, x.shape[-1]) + visual_temporal_embed.view(1, T, 1, x.shape[-1])
|
1226 |
+
|
1227 |
+
x_feat_dict = {}
|
1228 |
+
|
1229 |
+
spatial_avg_pool_x = x.view(batch_size, T, -1, x.shape[-1]).mean(dim=2)
|
1230 |
+
x_feat_dict['spatial_avg_pool'] = spatial_avg_pool_x
|
1231 |
+
|
1232 |
+
temporal_avg_pool_x = x.view(batch_size, T, -1, x.shape[-1]).mean(dim=1)
|
1233 |
+
x_feat_dict['temporal_avg_pool'] = temporal_avg_pool_x
|
1234 |
+
|
1235 |
+
x = x.view(batch_size, T, -1, x.shape[-1])[:, -1]
|
1236 |
+
x_feat_dict['last_frame'] = x
|
1237 |
+
|
1238 |
+
new_x = []
|
1239 |
+
for _image_feature_source in self.image_feature_source:
|
1240 |
+
if _image_feature_source not in x_feat_dict:
|
1241 |
+
raise ValueError('invalid image feature source: {}'.format(_image_feature_source))
|
1242 |
+
new_x.append(x_feat_dict[_image_feature_source])
|
1243 |
+
|
1244 |
+
x = torch.cat(new_x, dim=1)
|
1245 |
+
|
1246 |
+
x = x @ self.image_projection
|
1247 |
+
x = self.image_proj_norm(x)
|
1248 |
+
|
1249 |
+
return x
|
1250 |
+
#######################################################
|
1251 |
+
|
1252 |
+
def get_input_embeddings(self):
|
1253 |
+
return self.model.embed_tokens
|
1254 |
+
|
1255 |
+
def set_input_embeddings(self, value):
|
1256 |
+
self.model.embed_tokens = value
|
1257 |
+
|
1258 |
+
def get_output_embeddings(self):
|
1259 |
+
return self.lm_head
|
1260 |
+
|
1261 |
+
def set_output_embeddings(self, new_embeddings):
|
1262 |
+
self.lm_head = new_embeddings
|
1263 |
+
|
1264 |
+
def set_decoder(self, decoder):
|
1265 |
+
self.model = decoder
|
1266 |
+
|
1267 |
+
def get_decoder(self):
|
1268 |
+
return self.model
|
1269 |
+
|
1270 |
+
@add_start_docstrings_to_model_forward(FEYNMODEL_INPUTS_DOCSTRING)
|
1271 |
+
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
1272 |
+
def forward(
|
1273 |
+
self,
|
1274 |
+
input_ids: torch.LongTensor = None,
|
1275 |
+
pixel_values: Optional[torch.Tensor] = None,
|
1276 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1277 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1278 |
+
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
1279 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1280 |
+
labels: Optional[torch.LongTensor] = None,
|
1281 |
+
use_cache: Optional[bool] = None,
|
1282 |
+
output_attentions: Optional[bool] = None,
|
1283 |
+
output_hidden_states: Optional[bool] = None,
|
1284 |
+
return_dict: Optional[bool] = None,
|
1285 |
+
cache_position: Optional[torch.LongTensor] = None,
|
1286 |
+
**kwargs,
|
1287 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
1288 |
+
r"""
|
1289 |
+
Args:
|
1290 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1291 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
1292 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
1293 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
1294 |
+
|
1295 |
+
Returns:
|
1296 |
+
|
1297 |
+
Example:
|
1298 |
+
|
1299 |
+
```python
|
1300 |
+
>>> from transformers import AutoTokenizer, GemmaForCausalLM
|
1301 |
+
|
1302 |
+
>>> model = GemmaForCausalLM.from_pretrained("google/gemma-2-9b")
|
1303 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
|
1304 |
+
|
1305 |
+
>>> prompt = "What is your favorite condiment?"
|
1306 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
1307 |
+
|
1308 |
+
>>> # Generate
|
1309 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
1310 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
1311 |
+
"What is your favorite condiment?"
|
1312 |
+
```"""
|
1313 |
+
|
1314 |
+
|
1315 |
+
if self.training and self.config._attn_implementation != "eager":
|
1316 |
+
logger.warning_once(
|
1317 |
+
"It is strongly recommended to train FeynModel models with the `eager` attention implementation "
|
1318 |
+
f"instead of `{self.config._attn_implementation}`. Use `eager` with `AutoModelForCausalLM.from_pretrained('<path-to-checkpoint>', attn_implementation='eager')`."
|
1319 |
+
)
|
1320 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1321 |
+
output_hidden_states = (
|
1322 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1323 |
+
)
|
1324 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1325 |
+
|
1326 |
+
if pixel_values is not None:
|
1327 |
+
self.model.mode='vlm'
|
1328 |
+
|
1329 |
+
if input_ids is not None:
|
1330 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
1331 |
+
image_features = self._encode_image(pixel_values)
|
1332 |
+
inputs_embeds, causal_attention_mask = self._merge_input_ids_with_image_features(image_features, inputs_embeds )
|
1333 |
+
causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=2048)
|
1334 |
+
causal_attention_mask=causal_attention_mask.to(input_ids.device)
|
1335 |
+
self.__causal_attention_mask=causal_attention_mask
|
1336 |
+
|
1337 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
1338 |
+
if pixel_values is not None:
|
1339 |
+
outputs = self.model(
|
1340 |
+
input_ids=None,
|
1341 |
+
attention_mask=causal_attention_mask,
|
1342 |
+
position_ids=position_ids,
|
1343 |
+
past_key_values=past_key_values,
|
1344 |
+
inputs_embeds=inputs_embeds,
|
1345 |
+
use_cache=use_cache,
|
1346 |
+
output_attentions=output_attentions,
|
1347 |
+
output_hidden_states=output_hidden_states,
|
1348 |
+
return_dict=return_dict,
|
1349 |
+
cache_position=cache_position,
|
1350 |
+
causal_attention_mask=causal_attention_mask,
|
1351 |
+
)
|
1352 |
+
else:
|
1353 |
+
outputs = self.model(
|
1354 |
+
input_ids=input_ids,
|
1355 |
+
attention_mask=attention_mask,
|
1356 |
+
position_ids=position_ids,
|
1357 |
+
past_key_values=past_key_values,
|
1358 |
+
inputs_embeds=inputs_embeds,
|
1359 |
+
use_cache=use_cache,
|
1360 |
+
output_attentions=output_attentions,
|
1361 |
+
output_hidden_states=output_hidden_states,
|
1362 |
+
return_dict=return_dict,
|
1363 |
+
cache_position=cache_position,
|
1364 |
+
causal_attention_mask=self.__causal_attention_mask,
|
1365 |
+
)
|
1366 |
+
|
1367 |
+
|
1368 |
+
hidden_states = outputs[0]
|
1369 |
+
logits = self.lm_head(hidden_states)
|
1370 |
+
|
1371 |
+
if self.config.final_logit_softcapping is not None:
|
1372 |
+
logits = logits / self.config.final_logit_softcapping
|
1373 |
+
logits = torch.tanh(logits)
|
1374 |
+
logits = logits * self.config.final_logit_softcapping
|
1375 |
+
|
1376 |
+
|
1377 |
+
logits = logits.float()
|
1378 |
+
loss = None
|
1379 |
+
if labels is not None:
|
1380 |
+
# we are doing next-token prediction; shift prediction scores and input ids by one
|
1381 |
+
num_image_tokens = self.model.image_patch_tokens
|
1382 |
+
shifted_logits = logits[:, num_image_tokens:-1, :].contiguous()
|
1383 |
+
labels = labels[:, 1:].contiguous()
|
1384 |
+
loss_fct = CrossEntropyLoss()
|
1385 |
+
loss = loss_fct(shifted_logits.view(-1, self.config.vocab_size), labels.view(-1))
|
1386 |
+
|
1387 |
+
if not return_dict:
|
1388 |
+
|
1389 |
+
output = (logits,) + outputs[1:]
|
1390 |
+
return (loss,) + output if loss is not None else output
|
1391 |
+
|
1392 |
+
return CausalLMOutputWithPast(
|
1393 |
+
loss=loss,
|
1394 |
+
logits=logits,
|
1395 |
+
past_key_values=outputs.past_key_values,
|
1396 |
+
hidden_states=outputs.hidden_states,
|
1397 |
+
attentions=outputs.attentions,
|
1398 |
+
)
|
1399 |
+
|
1400 |
+
def prepare_inputs_for_generation(
|
1401 |
+
self,
|
1402 |
+
input_ids,
|
1403 |
+
past_key_values=None,
|
1404 |
+
attention_mask=None,
|
1405 |
+
inputs_embeds=None,
|
1406 |
+
cache_position=None,
|
1407 |
+
position_ids=None,
|
1408 |
+
use_cache=True,
|
1409 |
+
**kwargs,
|
1410 |
+
):
|
1411 |
+
|
1412 |
+
|
1413 |
+
# If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
|
1414 |
+
# Exception 1: when passing input_embeds, input_ids may be missing entries
|
1415 |
+
# Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
|
1416 |
+
if past_key_values is not None:
|
1417 |
+
if inputs_embeds is not None: # Exception 1
|
1418 |
+
input_ids = input_ids[:, -cache_position.shape[0] :]
|
1419 |
+
elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2)
|
1420 |
+
input_ids = input_ids[:, cache_position]
|
1421 |
+
|
1422 |
+
if attention_mask is not None and position_ids is None:
|
1423 |
+
# create position_ids on the fly for batch generation
|
1424 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
1425 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
1426 |
+
if past_key_values:
|
1427 |
+
# print(f"+-+-+-+-+-+-+++ past_key_values +-+-+++- position_ids {position_ids.size()} ================= ")
|
1428 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
1429 |
+
# This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s
|
1430 |
+
# `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride
|
1431 |
+
# during the decoding. Here, simply using `.contiguous()` is not sufficient as in the
|
1432 |
+
# batch size = 1 case, `position_ids` is already contiguous but with varying stride
|
1433 |
+
# which retriggers a capture.
|
1434 |
+
position_ids = position_ids.clone(memory_format=torch.contiguous_format)
|
1435 |
+
# print(f"+-+-+-+-+-+-+++ past_key_values +-+-+++- position_ids cmlone ==> {position_ids.size()} ================= ")
|
1436 |
+
|
1437 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
1438 |
+
if inputs_embeds is not None and cache_position[0] == 0:
|
1439 |
+
#print(">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> first generation step>>>>>>>>>>>>>>>>>>>>>>>>>>>>>><")
|
1440 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
1441 |
+
else:
|
1442 |
+
# The clone here is for the same reason as for `position_ids`.
|
1443 |
+
# print(">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> The clone here is for the same reason as for `position_ids` ==> input_ids input_ids.clone.>>>>>>>>>>>>>>>>>>>>>>>>>>>>>><")
|
1444 |
+
model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format)}
|
1445 |
+
|
1446 |
+
if isinstance(past_key_values, HybridCache) and attention_mask.ndim == 2:
|
1447 |
+
if inputs_embeds is not None and input_ids.size(1)!= 0 :
|
1448 |
+
###################### V ############## add _ for _ = inputs_embeds.shape
|
1449 |
+
batch_size, sequence_length, _ = inputs_embeds.shape
|
1450 |
+
device = inputs_embeds.device
|
1451 |
+
#print(f"1111111 +-+-+-+-+-+-+-+-+-+- sequence_length = inputs_embeds {sequence_length}")
|
1452 |
+
else:
|
1453 |
+
batch_size, sequence_length = position_ids.shape
|
1454 |
+
device = input_ids.device
|
1455 |
+
#print(f"22222222 +-+-+-+-+-+-+-+-+-+- sequence_length = input_ids.shape {sequence_length}")
|
1456 |
+
|
1457 |
+
dtype = self.lm_head.weight.dtype
|
1458 |
+
min_dtype = torch.finfo(dtype).min
|
1459 |
+
|
1460 |
+
attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
|
1461 |
+
attention_mask,
|
1462 |
+
sequence_length=sequence_length,
|
1463 |
+
target_length=past_key_values.get_max_length(),
|
1464 |
+
dtype=dtype,
|
1465 |
+
device=device,
|
1466 |
+
min_dtype=min_dtype,
|
1467 |
+
cache_position=cache_position,
|
1468 |
+
batch_size=batch_size,
|
1469 |
+
)
|
1470 |
+
|
1471 |
+
|
1472 |
+
model_inputs.update(
|
1473 |
+
{
|
1474 |
+
"position_ids": position_ids,
|
1475 |
+
"cache_position": cache_position,
|
1476 |
+
"past_key_values": past_key_values,
|
1477 |
+
"use_cache": use_cache,
|
1478 |
+
"attention_mask": attention_mask,
|
1479 |
+
}
|
1480 |
+
)
|
1481 |
+
return model_inputs
|
1482 |
+
|
1483 |
+
def generate(
|
1484 |
+
self,
|
1485 |
+
input_ids,
|
1486 |
+
pixel_values=None,
|
1487 |
+
max_length=None,
|
1488 |
+
do_sample=True,
|
1489 |
+
temperature=0.7,
|
1490 |
+
**kwargs
|
1491 |
+
):
|
1492 |
+
|
1493 |
+
|
1494 |
+
if pixel_values is not None:
|
1495 |
+
if input_ids is not None:
|
1496 |
+
|
1497 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
1498 |
+
print("pixels")
|
1499 |
+
image_features = self._encode_image(pixel_values)
|
1500 |
+
inputs_embeds, causal_attention_mask = self._merge_input_ids_with_image_features(image_features, inputs_embeds )
|
1501 |
+
causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=max_length)
|
1502 |
+
causal_attention_mask=causal_attention_mask.to(input_ids.device)
|
1503 |
+
self.__causal_attention_mask=causal_attention_mask
|
1504 |
+
self.model.mode='vlm'
|
1505 |
+
result = super().generate(
|
1506 |
+
input_ids=None,
|
1507 |
+
inputs_embeds=inputs_embeds,
|
1508 |
+
max_length=max_length,
|
1509 |
+
do_sample=do_sample,
|
1510 |
+
temperature=temperature,
|
1511 |
+
**kwargs
|
1512 |
+
)
|
1513 |
+
|
1514 |
+
else:
|
1515 |
+
|
1516 |
+
self.model.mode=='llm'
|
1517 |
+
result = super().generate(
|
1518 |
+
input_ids=input_ids,
|
1519 |
+
#inputs_embeds=None,
|
1520 |
+
max_length=max_length,
|
1521 |
+
do_sample=do_sample,
|
1522 |
+
temperature=temperature,
|
1523 |
+
**kwargs
|
1524 |
+
)
|
1525 |
+
self.__causal_attention_mask = None
|
1526 |
+
|
1527 |
+
return result
|
1528 |
+
|
preprocessor_config.json
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoProcessor": "processing_florence2.Florence2Processor"
|
4 |
+
},
|
5 |
+
"crop_size": {
|
6 |
+
"height": 768,
|
7 |
+
"width": 768
|
8 |
+
},
|
9 |
+
"do_center_crop": false,
|
10 |
+
"do_convert_rgb": null,
|
11 |
+
"do_normalize": true,
|
12 |
+
"do_rescale": true,
|
13 |
+
"do_resize": true,
|
14 |
+
"image_mean": [
|
15 |
+
0.485,
|
16 |
+
0.456,
|
17 |
+
0.406
|
18 |
+
],
|
19 |
+
"image_processor_type": "CLIPImageProcessor",
|
20 |
+
"image_seq_length": 577,
|
21 |
+
"image_std": [
|
22 |
+
0.229,
|
23 |
+
0.224,
|
24 |
+
0.225
|
25 |
+
],
|
26 |
+
"processor_class": "Florence2Processor",
|
27 |
+
"resample": 3,
|
28 |
+
"rescale_factor": 0.00392156862745098,
|
29 |
+
"size": {
|
30 |
+
"height": 768,
|
31 |
+
"width": 768
|
32 |
+
}
|
33 |
+
}
|
processing_florence2.py
ADDED
@@ -0,0 +1,1088 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2024 Microsoft and The HuggingFace Inc. team.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
"""
|
16 |
+
Processor class for Florence-2.
|
17 |
+
"""
|
18 |
+
|
19 |
+
import re
|
20 |
+
import logging
|
21 |
+
from typing import List, Optional, Union
|
22 |
+
import numpy as np
|
23 |
+
|
24 |
+
import torch
|
25 |
+
|
26 |
+
from transformers.feature_extraction_utils import BatchFeature
|
27 |
+
from transformers.image_utils import ImageInput, is_valid_image
|
28 |
+
from transformers.processing_utils import ProcessorMixin
|
29 |
+
from transformers.tokenization_utils_base import (
|
30 |
+
PaddingStrategy,
|
31 |
+
PreTokenizedInput,
|
32 |
+
TextInput,
|
33 |
+
TruncationStrategy,
|
34 |
+
)
|
35 |
+
from transformers.utils import TensorType
|
36 |
+
|
37 |
+
|
38 |
+
logger = logging.getLogger(__name__)
|
39 |
+
|
40 |
+
# Copied from transformers.models.idefics2.processing_idefics2.is_url
|
41 |
+
def is_url(val) -> bool:
|
42 |
+
return isinstance(val, str) and val.startswith("http")
|
43 |
+
|
44 |
+
# Copied from transformers.models.idefics2.processing_idefics2.is_image_or_image_url
|
45 |
+
def is_image_or_image_url(elem):
|
46 |
+
return is_url(elem) or is_valid_image(elem)
|
47 |
+
|
48 |
+
|
49 |
+
def _is_str_or_image(elem):
|
50 |
+
return isinstance(elem, (str)) or is_image_or_image_url(elem)
|
51 |
+
|
52 |
+
|
53 |
+
class Florence2Processor(ProcessorMixin):
|
54 |
+
r"""
|
55 |
+
Constructs a Florence2 processor which wraps a Florence2 image processor and a Florence2 tokenizer into a single processor.
|
56 |
+
|
57 |
+
[`Florence2Processor`] offers all the functionalities of [`CLIPImageProcessor`] and [`BartTokenizerFast`]. See the
|
58 |
+
[`~Florence2Processor.__call__`] and [`~Florence2Processor.decode`] for more information.
|
59 |
+
|
60 |
+
Args:
|
61 |
+
image_processor ([`CLIPImageProcessor`], *optional*):
|
62 |
+
The image processor is a required input.
|
63 |
+
tokenizer ([`BartTokenizerFast`], *optional*):
|
64 |
+
The tokenizer is a required input.
|
65 |
+
"""
|
66 |
+
|
67 |
+
attributes = ["image_processor", "tokenizer"]
|
68 |
+
image_processor_class = "CLIPImageProcessor"
|
69 |
+
tokenizer_class = ("BartTokenizer", "BartTokenizerFast")
|
70 |
+
|
71 |
+
def __init__(
|
72 |
+
self,
|
73 |
+
image_processor=None,
|
74 |
+
tokenizer=None,
|
75 |
+
):
|
76 |
+
if image_processor is None:
|
77 |
+
raise ValueError("You need to specify an `image_processor`.")
|
78 |
+
if tokenizer is None:
|
79 |
+
raise ValueError("You need to specify a `tokenizer`.")
|
80 |
+
if not hasattr(image_processor, "image_seq_length"):
|
81 |
+
raise ValueError("Image processor is missing an `image_seq_length` attribute.")
|
82 |
+
|
83 |
+
self.image_seq_length = image_processor.image_seq_length
|
84 |
+
|
85 |
+
tokens_to_add = {
|
86 |
+
'additional_special_tokens': \
|
87 |
+
tokenizer.additional_special_tokens + \
|
88 |
+
['<od>', '</od>', '<ocr>', '</ocr>'] + \
|
89 |
+
[f'<loc_{x}>' for x in range(1000)] + \
|
90 |
+
['<cap>', '</cap>', '<ncap>', '</ncap>','<dcap>', '</dcap>', '<grounding>', '</grounding>', '<seg>', '</seg>', '<sep>', '<region_cap>', '</region_cap>', '<region_to_desciption>', '</region_to_desciption>', '<proposal>', '</proposal>', '<poly>', '</poly>', '<and>']
|
91 |
+
}
|
92 |
+
tokenizer.add_special_tokens(tokens_to_add)
|
93 |
+
|
94 |
+
self.tasks_answer_post_processing_type = {
|
95 |
+
'<OCR>': 'pure_text',
|
96 |
+
'<OCR_WITH_REGION>': 'ocr',
|
97 |
+
'<CAPTION>': 'pure_text',
|
98 |
+
'<DETAILED_CAPTION>': 'pure_text',
|
99 |
+
'<MORE_DETAILED_CAPTION>': 'pure_text',
|
100 |
+
'<OD>': 'description_with_bboxes',
|
101 |
+
'<DENSE_REGION_CAPTION>': 'description_with_bboxes',
|
102 |
+
'<CAPTION_TO_PHRASE_GROUNDING>': "phrase_grounding",
|
103 |
+
'<REFERRING_EXPRESSION_SEGMENTATION>': 'polygons',
|
104 |
+
'<REGION_TO_SEGMENTATION>': 'polygons',
|
105 |
+
'<OPEN_VOCABULARY_DETECTION>': 'description_with_bboxes_or_polygons',
|
106 |
+
'<REGION_TO_CATEGORY>': 'pure_text',
|
107 |
+
'<REGION_TO_DESCRIPTION>': 'pure_text',
|
108 |
+
'<REGION_TO_OCR>': 'pure_text',
|
109 |
+
'<REGION_PROPOSAL>': 'bboxes'
|
110 |
+
}
|
111 |
+
|
112 |
+
self.task_prompts_without_inputs = {
|
113 |
+
'<OCR>': 'What is the text in the image?',
|
114 |
+
'<OCR_WITH_REGION>': 'What is the text in the image, with regions?',
|
115 |
+
'<CAPTION>': 'What does the image describe?',
|
116 |
+
'<DETAILED_CAPTION>': 'Describe in detail what is shown in the image.',
|
117 |
+
'<MORE_DETAILED_CAPTION>': 'Describe with a paragraph what is shown in the image.',
|
118 |
+
'<OD>': 'Locate the objects with category name in the image.',
|
119 |
+
'<DENSE_REGION_CAPTION>': 'Locate the objects in the image, with their descriptions.',
|
120 |
+
'<REGION_PROPOSAL>': 'Locate the region proposals in the image.'
|
121 |
+
}
|
122 |
+
|
123 |
+
self.task_prompts_with_input = {
|
124 |
+
'<CAPTION_TO_PHRASE_GROUNDING>': "Locate the phrases in the caption: {input}",
|
125 |
+
'<REFERRING_EXPRESSION_SEGMENTATION>': 'Locate {input} in the image with mask',
|
126 |
+
'<REGION_TO_SEGMENTATION>': 'What is the polygon mask of region {input}',
|
127 |
+
'<OPEN_VOCABULARY_DETECTION>': 'Locate {input} in the image.',
|
128 |
+
'<REGION_TO_CATEGORY>': 'What is the region {input}?',
|
129 |
+
'<REGION_TO_DESCRIPTION>': 'What does the region {input} describe?',
|
130 |
+
'<REGION_TO_OCR>': 'What text is in the region {input}?',
|
131 |
+
}
|
132 |
+
|
133 |
+
self.post_processor = Florence2PostProcesser(tokenizer=tokenizer)
|
134 |
+
|
135 |
+
|
136 |
+
super().__init__(image_processor, tokenizer)
|
137 |
+
|
138 |
+
def _construct_prompts(self, text):
|
139 |
+
# replace the task tokens with the task prompts if task token is in the text
|
140 |
+
prompts = []
|
141 |
+
for _text in text:
|
142 |
+
# 1. fixed task prompts without additional inputs
|
143 |
+
for task_token, task_prompt in self.task_prompts_without_inputs.items():
|
144 |
+
if task_token in _text:
|
145 |
+
assert _text == task_token, f"Task token {task_token} should be the only token in the text."
|
146 |
+
_text = task_prompt
|
147 |
+
break
|
148 |
+
# 2. task prompts with additional inputs
|
149 |
+
for task_token, task_prompt in self.task_prompts_with_input.items():
|
150 |
+
if task_token in _text:
|
151 |
+
_text = task_prompt.format(input=_text.replace(task_token, ''))
|
152 |
+
break
|
153 |
+
prompts.append(_text)
|
154 |
+
return prompts
|
155 |
+
|
156 |
+
def __call__(
|
157 |
+
self,
|
158 |
+
text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
|
159 |
+
images: ImageInput = None,
|
160 |
+
tokenize_newline_separately: bool = True,
|
161 |
+
padding: Union[bool, str, PaddingStrategy] = False,
|
162 |
+
truncation: Union[bool, str, TruncationStrategy] = None,
|
163 |
+
max_length=None,
|
164 |
+
return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
|
165 |
+
do_resize: bool = None,
|
166 |
+
do_normalize: bool = None,
|
167 |
+
image_mean: Optional[Union[float, List[float]]] = None,
|
168 |
+
image_std: Optional[Union[float, List[float]]] = None,
|
169 |
+
data_format: Optional["ChannelDimension"] = "channels_first", # noqa: F821
|
170 |
+
input_data_format: Optional[
|
171 |
+
Union[str, "ChannelDimension"] # noqa: F821
|
172 |
+
] = None,
|
173 |
+
resample: "PILImageResampling" = None, # noqa: F821
|
174 |
+
do_convert_rgb: bool = None,
|
175 |
+
do_thumbnail: bool = None,
|
176 |
+
do_align_long_axis: bool = None,
|
177 |
+
do_rescale: bool = None,
|
178 |
+
) -> BatchFeature:
|
179 |
+
"""
|
180 |
+
Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
|
181 |
+
and `kwargs` arguments to BartTokenizerFast's [`~BartTokenizerFast.__call__`] if `text` is not `None` to encode
|
182 |
+
the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
|
183 |
+
CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
|
184 |
+
of the above two methods for more information.
|
185 |
+
|
186 |
+
Args:
|
187 |
+
text (`str`, `List[str]`, `List[List[str]]`):
|
188 |
+
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
|
189 |
+
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
|
190 |
+
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
|
191 |
+
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
|
192 |
+
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
|
193 |
+
tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
|
194 |
+
number of channels, H and W are image height and width.
|
195 |
+
tokenize_newline_separately (`bool`, defaults to `True`):
|
196 |
+
Adds a separately tokenized '\n' at the end of the prompt.
|
197 |
+
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
|
198 |
+
Select a strategy to pad the returned sequences (according to the model's padding side and padding
|
199 |
+
index) among:
|
200 |
+
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
|
201 |
+
sequence if provided).
|
202 |
+
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
|
203 |
+
acceptable input length for the model if that argument is not provided.
|
204 |
+
- `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
|
205 |
+
lengths).
|
206 |
+
max_length (`int`, *optional*):
|
207 |
+
Maximum length of the returned list and optionally padding length (see above).
|
208 |
+
truncation (`bool`, *optional*):
|
209 |
+
Activates truncation to cut input sequences longer than `max_length` to `max_length`.
|
210 |
+
return_tensors (`str` or [`~utils.TensorType`], *optional*):
|
211 |
+
If set, will return tensors of a particular framework. Acceptable values are:
|
212 |
+
|
213 |
+
- `'tf'`: Return TensorFlow `tf.constant` objects.
|
214 |
+
- `'pt'`: Return PyTorch `torch.Tensor` objects.
|
215 |
+
- `'np'`: Return NumPy `np.ndarray` objects.
|
216 |
+
- `'jax'`: Return JAX `jnp.ndarray` objects.
|
217 |
+
|
218 |
+
Returns:
|
219 |
+
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
|
220 |
+
|
221 |
+
- **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`. If `suffix`
|
222 |
+
is provided, the `input_ids` will also contain the suffix input ids.
|
223 |
+
- **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
|
224 |
+
`return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
|
225 |
+
`None`).
|
226 |
+
- **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
|
227 |
+
- **labels** -- Labels compatible with training if `suffix` is not None
|
228 |
+
"""
|
229 |
+
|
230 |
+
return_token_type_ids = False
|
231 |
+
|
232 |
+
if images is None:
|
233 |
+
raise ValueError("`images` are expected as arguments to a `Florence2Processor` instance.")
|
234 |
+
if text is None:
|
235 |
+
logger.warning_once(
|
236 |
+
"You are using Florence-2 without a text prompt."
|
237 |
+
)
|
238 |
+
text = ""
|
239 |
+
|
240 |
+
if isinstance(text, List) and isinstance(images, List):
|
241 |
+
if len(images) < len(text):
|
242 |
+
raise ValueError(
|
243 |
+
f"Received {len(images)} images for {len(text)} prompts. Each prompt should be associated with an image."
|
244 |
+
)
|
245 |
+
if _is_str_or_image(text):
|
246 |
+
text = [text]
|
247 |
+
elif isinstance(text, list) and _is_str_or_image(text[0]):
|
248 |
+
pass
|
249 |
+
|
250 |
+
pixel_values = self.image_processor(
|
251 |
+
images,
|
252 |
+
do_resize=do_resize,
|
253 |
+
do_normalize=do_normalize,
|
254 |
+
return_tensors=return_tensors,
|
255 |
+
image_mean=image_mean,
|
256 |
+
image_std=image_std,
|
257 |
+
input_data_format=input_data_format,
|
258 |
+
data_format=data_format,
|
259 |
+
resample=resample,
|
260 |
+
do_convert_rgb=do_convert_rgb,
|
261 |
+
)["pixel_values"]
|
262 |
+
|
263 |
+
if max_length is not None:
|
264 |
+
max_length -= self.image_seq_length # max_length has to account for the image tokens
|
265 |
+
|
266 |
+
text = self._construct_prompts(text)
|
267 |
+
|
268 |
+
inputs = self.tokenizer(
|
269 |
+
text,
|
270 |
+
return_tensors=return_tensors,
|
271 |
+
padding=padding,
|
272 |
+
max_length=max_length,
|
273 |
+
truncation=truncation,
|
274 |
+
return_token_type_ids=return_token_type_ids,
|
275 |
+
)
|
276 |
+
|
277 |
+
return_data = {**inputs, "pixel_values": pixel_values}
|
278 |
+
|
279 |
+
if return_token_type_ids:
|
280 |
+
labels = inputs["input_ids"].masked_fill(inputs["token_type_ids"] == 0, -100)
|
281 |
+
return_data.update({"labels": labels})
|
282 |
+
return BatchFeature(data=return_data)
|
283 |
+
|
284 |
+
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Florence2
|
285 |
+
def batch_decode(self, *args, **kwargs):
|
286 |
+
"""
|
287 |
+
This method forwards all its arguments to BartTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
|
288 |
+
refer to the docstring of this method for more information.
|
289 |
+
"""
|
290 |
+
return self.tokenizer.batch_decode(*args, **kwargs)
|
291 |
+
|
292 |
+
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Florence2
|
293 |
+
def decode(self, *args, **kwargs):
|
294 |
+
"""
|
295 |
+
This method forwards all its arguments to BartTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
|
296 |
+
the docstring of this method for more information.
|
297 |
+
"""
|
298 |
+
return self.tokenizer.decode(*args, **kwargs)
|
299 |
+
|
300 |
+
@property
|
301 |
+
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->Florence2
|
302 |
+
def model_input_names(self):
|
303 |
+
tokenizer_input_names = self.tokenizer.model_input_names
|
304 |
+
image_processor_input_names = self.image_processor.model_input_names
|
305 |
+
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
306 |
+
|
307 |
+
def post_process_generation(self, text, task, image_size):
|
308 |
+
"""
|
309 |
+
Post-process the output of the model to each of the task outputs.
|
310 |
+
|
311 |
+
Args:
|
312 |
+
text (`str`): The text to post-process.
|
313 |
+
task (`str`): The task to post-process the text for.
|
314 |
+
image_size (`Tuple[int, int]`): The size of the image. height x width.
|
315 |
+
"""
|
316 |
+
|
317 |
+
task_answer_post_processing_type = self.tasks_answer_post_processing_type.get(task, 'pure_text')
|
318 |
+
task_answer = self.post_processor(
|
319 |
+
text=text,
|
320 |
+
image_size=image_size,
|
321 |
+
parse_tasks=task_answer_post_processing_type,
|
322 |
+
)[task_answer_post_processing_type]
|
323 |
+
|
324 |
+
if task_answer_post_processing_type == 'pure_text':
|
325 |
+
final_answer = task_answer
|
326 |
+
# remove the special tokens
|
327 |
+
final_answer = final_answer.replace('<s>', '').replace('</s>', '')
|
328 |
+
elif task_answer_post_processing_type in ['od', 'description_with_bboxes', 'bboxes']:
|
329 |
+
od_instances = task_answer
|
330 |
+
bboxes_od = [_od_instance['bbox'] for _od_instance in od_instances]
|
331 |
+
labels_od = [str(_od_instance['cat_name']) for _od_instance in od_instances]
|
332 |
+
final_answer = {'bboxes': bboxes_od, 'labels': labels_od}
|
333 |
+
elif task_answer_post_processing_type in ['ocr']:
|
334 |
+
bboxes = [_od_instance['quad_box'] for _od_instance in task_answer]
|
335 |
+
labels = [str(_od_instance['text']) for _od_instance in task_answer]
|
336 |
+
final_answer = {'quad_boxes': bboxes, 'labels': labels}
|
337 |
+
elif task_answer_post_processing_type in ['phrase_grounding']:
|
338 |
+
bboxes = []
|
339 |
+
labels = []
|
340 |
+
for _grounded_phrase in task_answer:
|
341 |
+
for _bbox in _grounded_phrase['bbox']:
|
342 |
+
bboxes.append(_bbox)
|
343 |
+
labels.append(_grounded_phrase['cat_name'])
|
344 |
+
final_answer = {'bboxes': bboxes, 'labels': labels}
|
345 |
+
elif task_answer_post_processing_type in ['description_with_polygons', 'polygons']:
|
346 |
+
labels = []
|
347 |
+
polygons = []
|
348 |
+
for result in task_answer:
|
349 |
+
label = result['cat_name']
|
350 |
+
_polygons = result['polygons']
|
351 |
+
labels.append(label)
|
352 |
+
polygons.append(_polygons)
|
353 |
+
final_answer = {'polygons': polygons, 'labels': labels}
|
354 |
+
elif task_answer_post_processing_type in ['description_with_bboxes_or_polygons']:
|
355 |
+
bboxes = []
|
356 |
+
bboxes_labels = []
|
357 |
+
polygons = []
|
358 |
+
polygons_labels = []
|
359 |
+
for result in task_answer:
|
360 |
+
label = result['cat_name']
|
361 |
+
if 'polygons' in result:
|
362 |
+
_polygons = result['polygons']
|
363 |
+
polygons.append(_polygons)
|
364 |
+
polygons_labels.append(label)
|
365 |
+
else:
|
366 |
+
_bbox = result['bbox']
|
367 |
+
bboxes.append(_bbox)
|
368 |
+
bboxes_labels.append(label)
|
369 |
+
final_answer = {'bboxes': bboxes, 'bboxes_labels': bboxes_labels, 'polygons': polygons, 'polygons_labels': polygons_labels}
|
370 |
+
else:
|
371 |
+
raise ValueError('Unknown task answer post processing type: {}'.format(task_answer_post_processing_type))
|
372 |
+
|
373 |
+
final_answer = {
|
374 |
+
task: final_answer}
|
375 |
+
return final_answer
|
376 |
+
|
377 |
+
class BoxQuantizer(object):
|
378 |
+
def __init__(self, mode, bins):
|
379 |
+
self.mode = mode
|
380 |
+
self.bins = bins
|
381 |
+
|
382 |
+
def quantize(self, boxes: torch.Tensor, size):
|
383 |
+
bins_w, bins_h = self.bins # Quantization bins.
|
384 |
+
size_w, size_h = size # Original image size.
|
385 |
+
size_per_bin_w = size_w / bins_w
|
386 |
+
size_per_bin_h = size_h / bins_h
|
387 |
+
xmin, ymin, xmax, ymax = boxes.split(1, dim=-1) # Shape: 4 * [N, 1].
|
388 |
+
|
389 |
+
if self.mode == 'floor':
|
390 |
+
quantized_xmin = (
|
391 |
+
xmin / size_per_bin_w).floor().clamp(0, bins_w - 1)
|
392 |
+
quantized_ymin = (
|
393 |
+
ymin / size_per_bin_h).floor().clamp(0, bins_h - 1)
|
394 |
+
quantized_xmax = (
|
395 |
+
xmax / size_per_bin_w).floor().clamp(0, bins_w - 1)
|
396 |
+
quantized_ymax = (
|
397 |
+
ymax / size_per_bin_h).floor().clamp(0, bins_h - 1)
|
398 |
+
|
399 |
+
elif self.mode == 'round':
|
400 |
+
raise NotImplementedError()
|
401 |
+
|
402 |
+
else:
|
403 |
+
raise ValueError('Incorrect quantization type.')
|
404 |
+
|
405 |
+
quantized_boxes = torch.cat(
|
406 |
+
(quantized_xmin, quantized_ymin, quantized_xmax, quantized_ymax), dim=-1
|
407 |
+
).int()
|
408 |
+
|
409 |
+
return quantized_boxes
|
410 |
+
|
411 |
+
def dequantize(self, boxes: torch.Tensor, size):
|
412 |
+
bins_w, bins_h = self.bins # Quantization bins.
|
413 |
+
size_w, size_h = size # Original image size.
|
414 |
+
size_per_bin_w = size_w / bins_w
|
415 |
+
size_per_bin_h = size_h / bins_h
|
416 |
+
xmin, ymin, xmax, ymax = boxes.split(1, dim=-1) # Shape: 4 * [N, 1].
|
417 |
+
|
418 |
+
if self.mode == 'floor':
|
419 |
+
# Add 0.5 to use the center position of the bin as the coordinate.
|
420 |
+
dequantized_xmin = (xmin + 0.5) * size_per_bin_w
|
421 |
+
dequantized_ymin = (ymin + 0.5) * size_per_bin_h
|
422 |
+
dequantized_xmax = (xmax + 0.5) * size_per_bin_w
|
423 |
+
dequantized_ymax = (ymax + 0.5) * size_per_bin_h
|
424 |
+
|
425 |
+
elif self.mode == 'round':
|
426 |
+
raise NotImplementedError()
|
427 |
+
|
428 |
+
else:
|
429 |
+
raise ValueError('Incorrect quantization type.')
|
430 |
+
|
431 |
+
dequantized_boxes = torch.cat(
|
432 |
+
(dequantized_xmin, dequantized_ymin,
|
433 |
+
dequantized_xmax, dequantized_ymax), dim=-1
|
434 |
+
)
|
435 |
+
|
436 |
+
return dequantized_boxes
|
437 |
+
|
438 |
+
|
439 |
+
class CoordinatesQuantizer(object):
|
440 |
+
"""
|
441 |
+
Quantize coornidates (Nx2)
|
442 |
+
"""
|
443 |
+
|
444 |
+
def __init__(self, mode, bins):
|
445 |
+
self.mode = mode
|
446 |
+
self.bins = bins
|
447 |
+
|
448 |
+
def quantize(self, coordinates: torch.Tensor, size):
|
449 |
+
bins_w, bins_h = self.bins # Quantization bins.
|
450 |
+
size_w, size_h = size # Original image size.
|
451 |
+
size_per_bin_w = size_w / bins_w
|
452 |
+
size_per_bin_h = size_h / bins_h
|
453 |
+
assert coordinates.shape[-1] == 2, 'coordinates should be shape (N, 2)'
|
454 |
+
x, y = coordinates.split(1, dim=-1) # Shape: 4 * [N, 1].
|
455 |
+
|
456 |
+
if self.mode == 'floor':
|
457 |
+
quantized_x = (x / size_per_bin_w).floor().clamp(0, bins_w - 1)
|
458 |
+
quantized_y = (y / size_per_bin_h).floor().clamp(0, bins_h - 1)
|
459 |
+
|
460 |
+
elif self.mode == 'round':
|
461 |
+
raise NotImplementedError()
|
462 |
+
|
463 |
+
else:
|
464 |
+
raise ValueError('Incorrect quantization type.')
|
465 |
+
|
466 |
+
quantized_coordinates = torch.cat(
|
467 |
+
(quantized_x, quantized_y), dim=-1
|
468 |
+
).int()
|
469 |
+
|
470 |
+
return quantized_coordinates
|
471 |
+
|
472 |
+
def dequantize(self, coordinates: torch.Tensor, size):
|
473 |
+
bins_w, bins_h = self.bins # Quantization bins.
|
474 |
+
size_w, size_h = size # Original image size.
|
475 |
+
size_per_bin_w = size_w / bins_w
|
476 |
+
size_per_bin_h = size_h / bins_h
|
477 |
+
assert coordinates.shape[-1] == 2, 'coordinates should be shape (N, 2)'
|
478 |
+
x, y = coordinates.split(1, dim=-1) # Shape: 4 * [N, 1].
|
479 |
+
|
480 |
+
if self.mode == 'floor':
|
481 |
+
# Add 0.5 to use the center position of the bin as the coordinate.
|
482 |
+
dequantized_x = (x + 0.5) * size_per_bin_w
|
483 |
+
dequantized_y = (y + 0.5) * size_per_bin_h
|
484 |
+
|
485 |
+
elif self.mode == 'round':
|
486 |
+
raise NotImplementedError()
|
487 |
+
|
488 |
+
else:
|
489 |
+
raise ValueError('Incorrect quantization type.')
|
490 |
+
|
491 |
+
dequantized_coordinates = torch.cat(
|
492 |
+
(dequantized_x, dequantized_y), dim=-1
|
493 |
+
)
|
494 |
+
|
495 |
+
return dequantized_coordinates
|
496 |
+
|
497 |
+
|
498 |
+
class Florence2PostProcesser(object):
|
499 |
+
"""
|
500 |
+
Florence-2 post process for converting text prediction to various tasks results.
|
501 |
+
|
502 |
+
Args:
|
503 |
+
config: A dict of configs.
|
504 |
+
tokenizer: A tokenizer for decoding text to spans.
|
505 |
+
sample config:
|
506 |
+
UNIFIED_POST_PROCESS:
|
507 |
+
# commom configs
|
508 |
+
NUM_BBOX_HEIGHT_BINS: 1000
|
509 |
+
NUM_BBOX_WIDTH_BINS: 1000
|
510 |
+
COORDINATES_HEIGHT_BINS: 1000
|
511 |
+
COORDINATES_WIDTH_BINS: 1000
|
512 |
+
# task specific configs, override the common configs
|
513 |
+
PRASE_TASKS:
|
514 |
+
- TASK_NAME: 'video_dense_caption'
|
515 |
+
PATTERN: 'r<time_(\d+)><time_(\d+)>([a-zA-Z0-9 ]+)'
|
516 |
+
SCORE_MODE: 'avg_cat_name_scores'
|
517 |
+
NUM_BINS: 100
|
518 |
+
- TASK_NAME: 'od'
|
519 |
+
PATTERN: 'r<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>([a-zA-Z0-9 ]+)'
|
520 |
+
SCORE_MODE: 'avg_cat_name_scores'
|
521 |
+
|
522 |
+
Returns:
|
523 |
+
parsed_dict (dict): A dict of parsed results.
|
524 |
+
"""
|
525 |
+
def __init__(
|
526 |
+
self,
|
527 |
+
tokenizer=None
|
528 |
+
):
|
529 |
+
parse_tasks = []
|
530 |
+
parse_task_configs = {}
|
531 |
+
config = self._create_default_config()
|
532 |
+
for task in config['PARSE_TASKS']:
|
533 |
+
parse_tasks.append(task['TASK_NAME'])
|
534 |
+
parse_task_configs[task['TASK_NAME']] = task
|
535 |
+
|
536 |
+
self.config = config
|
537 |
+
self.parse_tasks = parse_tasks
|
538 |
+
self.parse_tasks_configs = parse_task_configs
|
539 |
+
|
540 |
+
self.tokenizer = tokenizer
|
541 |
+
if self.tokenizer is not None:
|
542 |
+
self.all_special_tokens = set(self.tokenizer.all_special_tokens)
|
543 |
+
|
544 |
+
self.init_quantizers()
|
545 |
+
self.black_list_of_phrase_grounding = self._create_black_list_of_phrase_grounding()
|
546 |
+
|
547 |
+
def _create_black_list_of_phrase_grounding(self):
|
548 |
+
black_list = {}
|
549 |
+
|
550 |
+
if 'phrase_grounding' in self.parse_tasks and self.parse_tasks_configs['phrase_grounding']['FILTER_BY_BLACK_LIST']:
|
551 |
+
black_list = set(
|
552 |
+
['it', 'I', 'me', 'mine',
|
553 |
+
'you', 'your', 'yours',
|
554 |
+
'he', 'him', 'his',
|
555 |
+
'she', 'her', 'hers',
|
556 |
+
'they', 'them', 'their', 'theirs',
|
557 |
+
'one', 'oneself',
|
558 |
+
'we', 'us', 'our', 'ours',
|
559 |
+
'you', 'your', 'yours',
|
560 |
+
'they', 'them', 'their', 'theirs',
|
561 |
+
'mine', 'yours', 'his', 'hers', 'its',
|
562 |
+
'ours', 'yours', 'theirs',
|
563 |
+
'myself', 'yourself', 'himself', 'herself', 'itself',
|
564 |
+
'ourselves', 'yourselves', 'themselves',
|
565 |
+
'this', 'that',
|
566 |
+
'these', 'those',
|
567 |
+
'who', 'whom', 'whose', 'which', 'what',
|
568 |
+
'who', 'whom', 'whose', 'which', 'that',
|
569 |
+
'all', 'another', 'any', 'anybody', 'anyone', 'anything',
|
570 |
+
'each', 'everybody', 'everyone', 'everything',
|
571 |
+
'few', 'many', 'nobody', 'none', 'one', 'several',
|
572 |
+
'some', 'somebody', 'someone', 'something',
|
573 |
+
'each other', 'one another',
|
574 |
+
'myself', 'yourself', 'himself', 'herself', 'itself',
|
575 |
+
'ourselves', 'yourselves', 'themselves',
|
576 |
+
'the image', 'image', 'images', 'the', 'a', 'an', 'a group',
|
577 |
+
'other objects', 'lots', 'a set',
|
578 |
+
]
|
579 |
+
)
|
580 |
+
|
581 |
+
return black_list
|
582 |
+
|
583 |
+
def _create_default_config(self):
|
584 |
+
config = {
|
585 |
+
'NUM_BBOX_HEIGHT_BINS': 1000,
|
586 |
+
'NUM_BBOX_WIDTH_BINS': 1000,
|
587 |
+
'BOX_QUANTIZATION_MODE': 'floor',
|
588 |
+
'COORDINATES_HEIGHT_BINS': 1000,
|
589 |
+
'COORDINATES_WIDTH_BINS': 1000,
|
590 |
+
'COORDINATES_QUANTIZATION_MODE': 'floor',
|
591 |
+
'PARSE_TASKS': [
|
592 |
+
{
|
593 |
+
'TASK_NAME': 'od',
|
594 |
+
'PATTERN': r'([a-zA-Z0-9 ]+)<loc_(\\d+)><loc_(\\d+)><loc_(\\d+)><loc_(\\d+)>'
|
595 |
+
},
|
596 |
+
{
|
597 |
+
'TASK_NAME': 'ocr',
|
598 |
+
'PATTERN': r'(.+?)<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>',
|
599 |
+
'AREA_THRESHOLD': 0.00
|
600 |
+
},
|
601 |
+
{
|
602 |
+
'TASK_NAME': 'phrase_grounding',
|
603 |
+
'FILTER_BY_BLACK_LIST': True
|
604 |
+
},
|
605 |
+
{
|
606 |
+
'TASK_NAME': 'pure_text',
|
607 |
+
},
|
608 |
+
{
|
609 |
+
'TASK_NAME': 'description_with_bboxes',
|
610 |
+
},
|
611 |
+
{
|
612 |
+
'TASK_NAME': 'description_with_polygons',
|
613 |
+
},
|
614 |
+
{
|
615 |
+
'TASK_NAME': 'polygons',
|
616 |
+
},
|
617 |
+
{
|
618 |
+
'TASK_NAME': 'bboxes',
|
619 |
+
},
|
620 |
+
{
|
621 |
+
'TASK_NAME': 'description_with_bboxes_or_polygons',
|
622 |
+
}
|
623 |
+
]
|
624 |
+
}
|
625 |
+
|
626 |
+
return config
|
627 |
+
|
628 |
+
def init_quantizers(self):
|
629 |
+
# we have box_quantizer (od, grounding) and coordinates_quantizer (ocr, referring_segmentation)
|
630 |
+
num_bbox_height_bins = self.config.get('NUM_BBOX_HEIGHT_BINS', 1000)
|
631 |
+
num_bbox_width_bins = self.config.get('NUM_BBOX_WIDTH_BINS', 1000)
|
632 |
+
box_quantization_mode = self.config.get('BOX_QUANTIZATION_MODE', 'floor')
|
633 |
+
self.box_quantizer = BoxQuantizer(
|
634 |
+
box_quantization_mode,
|
635 |
+
(num_bbox_width_bins, num_bbox_height_bins),
|
636 |
+
)
|
637 |
+
|
638 |
+
num_bbox_height_bins = self.config['COORDINATES_HEIGHT_BINS'] if 'COORDINATES_HEIGHT_BINS' in self.config else self.config.get('NUM_BBOX_HEIGHT_BINS', 1000)
|
639 |
+
num_bbox_width_bins = self.config['COORDINATES_WIDTH_BINS'] if 'COORDINATES_WIDTH_BINS' in self.config else self.config.get('NUM_BBOX_WIDTH_BINS', 1000)
|
640 |
+
box_quantization_mode = self.config.get('COORDINATES_QUANTIZATION_MODE') if 'COORDINATES_QUANTIZATION_MODE' in self.config else self.config.get('BOX_QUANTIZATION_MODE', 'floor')
|
641 |
+
self.coordinates_quantizer = CoordinatesQuantizer(
|
642 |
+
box_quantization_mode,
|
643 |
+
(num_bbox_width_bins, num_bbox_height_bins),
|
644 |
+
)
|
645 |
+
|
646 |
+
def decode_with_spans(self, tokenizer, token_ids):
|
647 |
+
filtered_tokens = tokenizer.convert_ids_to_tokens(
|
648 |
+
token_ids, skip_special_tokens=False)
|
649 |
+
assert len(filtered_tokens) == len(token_ids)
|
650 |
+
|
651 |
+
# To avoid mixing byte-level and unicode for byte-level BPT
|
652 |
+
# we need to build string separately for added tokens and byte-level tokens
|
653 |
+
# cf. https://github.com/huggingface/transformers/issues/1133
|
654 |
+
sub_texts = []
|
655 |
+
for token in filtered_tokens:
|
656 |
+
if token in self.all_special_tokens:
|
657 |
+
sub_texts.append(token)
|
658 |
+
else:
|
659 |
+
if isinstance(tokenizer, (BartTokenizer, BartTokenizerFast)):
|
660 |
+
sub_text = tokenizer.convert_tokens_to_string([token])
|
661 |
+
elif isinstance(tokenizer, (T5Tokenizer, T5TokenizerFast)):
|
662 |
+
# Ref: https://github.com/google/sentencepiece#whitespace-is-treated-as-a-basic-symbol
|
663 |
+
# Note: Do not strip sub_text as it may have functional whitespace
|
664 |
+
sub_text = token.replace('▁', ' ')
|
665 |
+
else:
|
666 |
+
raise ValueError(f'type {type(tokenizer)} not supported')
|
667 |
+
sub_texts.append(sub_text)
|
668 |
+
|
669 |
+
text = ''
|
670 |
+
spans = []
|
671 |
+
for sub_text in sub_texts:
|
672 |
+
span = (len(text), len(text) + len(sub_text)) # [start index, end index).
|
673 |
+
text += sub_text
|
674 |
+
spans.append(span)
|
675 |
+
|
676 |
+
# Text format:
|
677 |
+
# 1. T5Tokenizer/T5TokenizerFast:
|
678 |
+
# "<loc_1><loc_2><loc_3><loc_4> transplanting dog<loc_1><loc_2><loc_3><loc_4> cat</s>"
|
679 |
+
# Equivalent to t5_tokenizer.decode(input_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False, spaces_between_special_tokens=False)
|
680 |
+
# 2. BartTokenizer (need to double check):
|
681 |
+
# "<s><loc_1><loc_2><loc_3><loc_4>transplanting dog<loc_1><loc_2><loc_3><loc_4>cat</s>"
|
682 |
+
# Equivalent to bart_tokenizer.decode(input_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False, spaces_between_special_tokens=False)
|
683 |
+
return text, spans
|
684 |
+
|
685 |
+
def parse_od_from_text_and_spans(
|
686 |
+
self,
|
687 |
+
text,
|
688 |
+
pattern,
|
689 |
+
image_size,
|
690 |
+
phrase_centric=False
|
691 |
+
):
|
692 |
+
parsed = list(re.finditer(pattern, text))
|
693 |
+
|
694 |
+
instances = []
|
695 |
+
for i in range(len(parsed)):
|
696 |
+
# Prepare instance.
|
697 |
+
instance = {}
|
698 |
+
|
699 |
+
if phrase_centric:
|
700 |
+
bbox_bins = [int(parsed[i].group(j)) for j in range(2, 6)]
|
701 |
+
else:
|
702 |
+
bbox_bins = [int(parsed[i].group(j)) for j in range(1, 5)]
|
703 |
+
instance['bbox'] = self.box_quantizer.dequantize(
|
704 |
+
boxes=torch.tensor(bbox_bins),
|
705 |
+
size=image_size
|
706 |
+
).tolist()
|
707 |
+
|
708 |
+
if phrase_centric:
|
709 |
+
instance['cat_name'] = parsed[i].group(1).lower().strip()
|
710 |
+
else:
|
711 |
+
instance['cat_name'] = parsed[i].group(5).lower().strip()
|
712 |
+
instances.append(instance)
|
713 |
+
|
714 |
+
return instances
|
715 |
+
|
716 |
+
def parse_ocr_from_text_and_spans(self,
|
717 |
+
text,
|
718 |
+
pattern,
|
719 |
+
image_size,
|
720 |
+
area_threshold=-1.0,
|
721 |
+
):
|
722 |
+
bboxes = []
|
723 |
+
labels = []
|
724 |
+
text = text.replace('<s>', '')
|
725 |
+
# ocr with regions
|
726 |
+
parsed = re.findall(pattern, text)
|
727 |
+
instances = []
|
728 |
+
image_width, image_height = image_size
|
729 |
+
|
730 |
+
for ocr_line in parsed:
|
731 |
+
ocr_content = ocr_line[0]
|
732 |
+
quad_box = ocr_line[1:]
|
733 |
+
quad_box = [int(i) for i in quad_box]
|
734 |
+
quad_box = self.coordinates_quantizer.dequantize(
|
735 |
+
torch.tensor(np.array(quad_box).reshape(-1, 2)),
|
736 |
+
size=image_size
|
737 |
+
).reshape(-1).tolist()
|
738 |
+
|
739 |
+
if area_threshold > 0:
|
740 |
+
x_coords = [i for i in quad_box[0::2]]
|
741 |
+
y_coords = [i for i in quad_box[1::2]]
|
742 |
+
|
743 |
+
# apply the Shoelace formula
|
744 |
+
area = 0.5 * abs(sum(x_coords[i] * y_coords[i + 1] - x_coords[i + 1] * y_coords[i] for i in range(4 - 1)))
|
745 |
+
|
746 |
+
if area < (image_width * image_height) * area_threshold:
|
747 |
+
continue
|
748 |
+
|
749 |
+
bboxes.append(quad_box)
|
750 |
+
labels.append(ocr_content)
|
751 |
+
instances.append({
|
752 |
+
'quad_box': quad_box,
|
753 |
+
'text': ocr_content,
|
754 |
+
})
|
755 |
+
return instances
|
756 |
+
|
757 |
+
def parse_phrase_grounding_from_text_and_spans(self, text, pattern, image_size):
|
758 |
+
# ignore <s> </s> and <pad>
|
759 |
+
cur_span = 0
|
760 |
+
if text.startswith('<s>'):
|
761 |
+
cur_span += 3
|
762 |
+
|
763 |
+
text = text.replace('<s>', '')
|
764 |
+
text = text.replace('</s>', '')
|
765 |
+
text = text.replace('<pad>', '')
|
766 |
+
|
767 |
+
pattern = r"([^<]+(?:<loc_\d+>){4,})"
|
768 |
+
phrases = re.findall(pattern, text)
|
769 |
+
|
770 |
+
# pattern should be text pattern and od pattern
|
771 |
+
pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_)'
|
772 |
+
box_pattern = r'<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>'
|
773 |
+
|
774 |
+
instances = []
|
775 |
+
for pharse_text in phrases:
|
776 |
+
phrase_text_strip = pharse_text.replace('<ground>', '', 1)
|
777 |
+
phrase_text_strip = pharse_text.replace('<obj>', '', 1)
|
778 |
+
|
779 |
+
if phrase_text_strip == '':
|
780 |
+
cur_span += len(pharse_text)
|
781 |
+
continue
|
782 |
+
|
783 |
+
# Prepare instance.
|
784 |
+
instance = {}
|
785 |
+
|
786 |
+
# parse phrase, get string
|
787 |
+
phrase = re.search(pattern, phrase_text_strip)
|
788 |
+
if phrase is None:
|
789 |
+
cur_span += len(pharse_text)
|
790 |
+
continue
|
791 |
+
|
792 |
+
# parse bboxes by box_pattern
|
793 |
+
bboxes_parsed = list(re.finditer(box_pattern, pharse_text))
|
794 |
+
if len(bboxes_parsed) == 0:
|
795 |
+
cur_span += len(pharse_text)
|
796 |
+
continue
|
797 |
+
|
798 |
+
phrase = phrase.group()
|
799 |
+
# remove leading and trailing spaces
|
800 |
+
phrase = phrase.strip()
|
801 |
+
|
802 |
+
if phrase in self.black_list_of_phrase_grounding:
|
803 |
+
cur_span += len(pharse_text)
|
804 |
+
continue
|
805 |
+
|
806 |
+
# a list of list
|
807 |
+
bbox_bins = [[int(_bboxes_parsed.group(j)) for j in range(1, 5)] for _bboxes_parsed in bboxes_parsed]
|
808 |
+
instance['bbox'] = self.box_quantizer.dequantize(
|
809 |
+
boxes=torch.tensor(bbox_bins),
|
810 |
+
size=image_size
|
811 |
+
).tolist()
|
812 |
+
|
813 |
+
# exclude non-ascii characters
|
814 |
+
phrase = phrase.encode('ascii',errors='ignore').decode('ascii')
|
815 |
+
instance['cat_name'] = phrase
|
816 |
+
|
817 |
+
instances.append(instance)
|
818 |
+
|
819 |
+
return instances
|
820 |
+
|
821 |
+
def parse_description_with_bboxes_from_text_and_spans(self, text, pattern, image_size, allow_empty_phrase=False):
|
822 |
+
# temporary parse solution, split by '.'
|
823 |
+
# ignore <s> </s> and <pad>
|
824 |
+
|
825 |
+
text = text.replace('<s>', '')
|
826 |
+
text = text.replace('</s>', '')
|
827 |
+
text = text.replace('<pad>', '')
|
828 |
+
|
829 |
+
if allow_empty_phrase:
|
830 |
+
pattern = rf"(?:(?:<loc_\d+>){{4,}})"
|
831 |
+
else:
|
832 |
+
pattern = r"([^<]+(?:<loc_\d+>){4,})"
|
833 |
+
phrases = re.findall(pattern, text)
|
834 |
+
|
835 |
+
# pattern should be text pattern and od pattern
|
836 |
+
pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_)'
|
837 |
+
box_pattern = r'<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>'
|
838 |
+
|
839 |
+
instances = []
|
840 |
+
for pharse_text in phrases:
|
841 |
+
phrase_text_strip = pharse_text.replace('<ground>', '', 1)
|
842 |
+
phrase_text_strip = pharse_text.replace('<obj>', '', 1)
|
843 |
+
|
844 |
+
if phrase_text_strip == '' and not allow_empty_phrase:
|
845 |
+
continue
|
846 |
+
|
847 |
+
# parse phrase, get string
|
848 |
+
phrase = re.search(pattern, phrase_text_strip)
|
849 |
+
if phrase is None:
|
850 |
+
continue
|
851 |
+
|
852 |
+
phrase = phrase.group()
|
853 |
+
# remove leading and trailing spaces
|
854 |
+
phrase = phrase.strip()
|
855 |
+
|
856 |
+
# parse bboxes by box_pattern
|
857 |
+
bboxes_parsed = list(re.finditer(box_pattern, pharse_text))
|
858 |
+
if len(bboxes_parsed) == 0:
|
859 |
+
continue
|
860 |
+
|
861 |
+
# a list of list
|
862 |
+
bbox_bins = [[int(_bboxes_parsed.group(j)) for j in range(1, 5)] for _bboxes_parsed in bboxes_parsed]
|
863 |
+
|
864 |
+
bboxes = self.box_quantizer.dequantize(
|
865 |
+
boxes=torch.tensor(bbox_bins),
|
866 |
+
size=image_size
|
867 |
+
).tolist()
|
868 |
+
|
869 |
+
phrase = phrase.encode('ascii',errors='ignore').decode('ascii')
|
870 |
+
for _bboxes in bboxes:
|
871 |
+
# Prepare instance.
|
872 |
+
instance = {}
|
873 |
+
instance['bbox'] = _bboxes
|
874 |
+
# exclude non-ascii characters
|
875 |
+
instance['cat_name'] = phrase
|
876 |
+
instances.append(instance)
|
877 |
+
|
878 |
+
return instances
|
879 |
+
|
880 |
+
def parse_description_with_polygons_from_text_and_spans(self, text, pattern, image_size,
|
881 |
+
allow_empty_phrase=False,
|
882 |
+
polygon_sep_token='<sep>',
|
883 |
+
polygon_start_token='<poly>',
|
884 |
+
polygon_end_token='</poly>',
|
885 |
+
with_box_at_start=False,
|
886 |
+
):
|
887 |
+
|
888 |
+
# ref_seg format: '<expression><x1><y1><x2><y2><><><sep><><><><>'
|
889 |
+
# ignore <s> </s> and <pad>
|
890 |
+
|
891 |
+
text = text.replace('<s>', '')
|
892 |
+
text = text.replace('</s>', '')
|
893 |
+
text = text.replace('<pad>', '')
|
894 |
+
|
895 |
+
if allow_empty_phrase:
|
896 |
+
pattern = rf"(?:(?:<loc_\d+>|{re.escape(polygon_sep_token)}|{re.escape(polygon_start_token)}|{re.escape(polygon_end_token)}){{4,}})"
|
897 |
+
else:
|
898 |
+
# [^<]+: This part matches one or more characters that are not the < symbol.
|
899 |
+
# The ^ inside the square brackets [] is a negation, meaning it matches anything except <.
|
900 |
+
#
|
901 |
+
pattern = rf"([^<]+(?:<loc_\d+>|{re.escape(polygon_sep_token)}|{re.escape(polygon_start_token)}|{re.escape(polygon_end_token)}){{4,}})"
|
902 |
+
phrases = re.findall(pattern, text)
|
903 |
+
|
904 |
+
phrase_string_pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_|<poly>)'
|
905 |
+
box_pattern = rf'((?:<loc_\d+>)+)(?:{re.escape(polygon_sep_token)}|$)'
|
906 |
+
|
907 |
+
# one polygons instance is separated by polygon_start_token and polygon_end_token
|
908 |
+
polygons_instance_pattern = rf'{re.escape(polygon_start_token)}(.*?){re.escape(polygon_end_token)}'
|
909 |
+
|
910 |
+
instances = []
|
911 |
+
for phrase_text in phrases:
|
912 |
+
|
913 |
+
# exclude loc_\d+>
|
914 |
+
# need to get span if want to include category score
|
915 |
+
phrase_text_strip = re.sub(r'^loc_\d+>', '', phrase_text, count=1)
|
916 |
+
|
917 |
+
# phrase = phrase.replace('<poly>', '')
|
918 |
+
# phrase = phrase.replace('poly>', '')
|
919 |
+
|
920 |
+
if phrase_text_strip == '' and not allow_empty_phrase:
|
921 |
+
continue
|
922 |
+
|
923 |
+
|
924 |
+
# parse phrase, get string
|
925 |
+
phrase = re.search(phrase_string_pattern, phrase_text_strip)
|
926 |
+
if phrase is None:
|
927 |
+
continue
|
928 |
+
phrase = phrase.group()
|
929 |
+
# remove leading and trailing spaces
|
930 |
+
phrase = phrase.strip()
|
931 |
+
|
932 |
+
# parse bboxes by box_pattern
|
933 |
+
|
934 |
+
# split by polygon_start_token and polygon_end_token first using polygons_instance_pattern
|
935 |
+
if polygon_start_token in phrase_text and polygon_end_token in phrase_text:
|
936 |
+
polygons_instances_parsed = list(re.finditer(polygons_instance_pattern, phrase_text))
|
937 |
+
else:
|
938 |
+
polygons_instances_parsed = [phrase_text]
|
939 |
+
|
940 |
+
for _polygons_instances_parsed in polygons_instances_parsed:
|
941 |
+
# Prepare instance.
|
942 |
+
instance = {}
|
943 |
+
|
944 |
+
# polygons_parsed= list(re.finditer(box_pattern, phrase_text))
|
945 |
+
if isinstance(_polygons_instances_parsed, str):
|
946 |
+
polygons_parsed= list(re.finditer(box_pattern, _polygons_instances_parsed))
|
947 |
+
else:
|
948 |
+
polygons_parsed= list(re.finditer(box_pattern, _polygons_instances_parsed.group(1)))
|
949 |
+
if len(polygons_parsed) == 0:
|
950 |
+
continue
|
951 |
+
|
952 |
+
# a list of list (polygon)
|
953 |
+
bbox = []
|
954 |
+
polygons = []
|
955 |
+
for _polygon_parsed in polygons_parsed:
|
956 |
+
# group 1: whole <loc_\d+>...</loc_\d+>
|
957 |
+
_polygon = _polygon_parsed.group(1)
|
958 |
+
# parse into list of int
|
959 |
+
_polygon = [int(_loc_parsed.group(1)) for _loc_parsed in re.finditer(r'<loc_(\d+)>', _polygon)]
|
960 |
+
if with_box_at_start and len(bbox) == 0:
|
961 |
+
if len(_polygon) > 4:
|
962 |
+
# no valid bbox prediction
|
963 |
+
bbox = _polygon[:4]
|
964 |
+
_polygon = _polygon[4:]
|
965 |
+
else:
|
966 |
+
bbox = [0, 0, 0, 0]
|
967 |
+
# abandon last element if is not paired
|
968 |
+
if len(_polygon) % 2 == 1:
|
969 |
+
_polygon = _polygon[:-1]
|
970 |
+
|
971 |
+
# reshape into (n, 2)
|
972 |
+
_polygon = self.coordinates_quantizer.dequantize(
|
973 |
+
torch.tensor(np.array(_polygon).reshape(-1, 2)),
|
974 |
+
size=image_size
|
975 |
+
).reshape(-1).tolist()
|
976 |
+
# reshape back
|
977 |
+
polygons.append(_polygon)
|
978 |
+
|
979 |
+
instance['cat_name'] = phrase
|
980 |
+
instance['polygons'] = polygons
|
981 |
+
if len(bbox) != 0:
|
982 |
+
instance['bbox'] = self.box_quantizer.dequantize(
|
983 |
+
boxes=torch.tensor([bbox]),
|
984 |
+
size=image_size
|
985 |
+
).tolist()[0]
|
986 |
+
|
987 |
+
instances.append(instance)
|
988 |
+
|
989 |
+
return instances
|
990 |
+
|
991 |
+
def __call__(
|
992 |
+
self,
|
993 |
+
text=None,
|
994 |
+
image_size=None,
|
995 |
+
parse_tasks=None,
|
996 |
+
):
|
997 |
+
"""
|
998 |
+
Args:
|
999 |
+
text: model outputs
|
1000 |
+
image_size: (width, height)
|
1001 |
+
parse_tasks: a list of tasks to parse, if None, parse all tasks.
|
1002 |
+
|
1003 |
+
"""
|
1004 |
+
if parse_tasks is not None:
|
1005 |
+
if isinstance(parse_tasks, str):
|
1006 |
+
parse_tasks = [parse_tasks]
|
1007 |
+
for _parse_task in parse_tasks:
|
1008 |
+
assert _parse_task in self.parse_tasks, f'parse task {_parse_task} not supported'
|
1009 |
+
|
1010 |
+
# sequence or text should be provided
|
1011 |
+
assert text is not None, 'text should be provided'
|
1012 |
+
|
1013 |
+
parsed_dict = {
|
1014 |
+
'text': text
|
1015 |
+
}
|
1016 |
+
|
1017 |
+
for task in self.parse_tasks:
|
1018 |
+
if parse_tasks is not None and task not in parse_tasks:
|
1019 |
+
continue
|
1020 |
+
|
1021 |
+
pattern = self.parse_tasks_configs[task].get('PATTERN', None)
|
1022 |
+
|
1023 |
+
if task == 'ocr':
|
1024 |
+
instances = self.parse_ocr_from_text_and_spans(
|
1025 |
+
text,
|
1026 |
+
pattern=pattern,
|
1027 |
+
image_size=image_size,
|
1028 |
+
area_threshold=self.parse_tasks_configs[task].get('AREA_THRESHOLD', 0.0),
|
1029 |
+
)
|
1030 |
+
parsed_dict['ocr'] = instances
|
1031 |
+
elif task == 'phrase_grounding':
|
1032 |
+
instances = self.parse_phrase_grounding_from_text_and_spans(
|
1033 |
+
text,
|
1034 |
+
pattern=pattern,
|
1035 |
+
image_size=image_size,
|
1036 |
+
)
|
1037 |
+
parsed_dict['phrase_grounding'] = instances
|
1038 |
+
elif task == 'pure_text':
|
1039 |
+
parsed_dict['pure_text'] = text
|
1040 |
+
elif task == 'description_with_bboxes':
|
1041 |
+
instances = self.parse_description_with_bboxes_from_text_and_spans(
|
1042 |
+
text,
|
1043 |
+
pattern=pattern,
|
1044 |
+
image_size=image_size,
|
1045 |
+
)
|
1046 |
+
parsed_dict['description_with_bboxes'] = instances
|
1047 |
+
elif task == 'description_with_polygons':
|
1048 |
+
instances = self.parse_description_with_polygons_from_text_and_spans(
|
1049 |
+
text,
|
1050 |
+
pattern=pattern,
|
1051 |
+
image_size=image_size,
|
1052 |
+
)
|
1053 |
+
parsed_dict['description_with_polygons'] = instances
|
1054 |
+
elif task == 'polygons':
|
1055 |
+
instances = self.parse_description_with_polygons_from_text_and_spans(
|
1056 |
+
text,
|
1057 |
+
pattern=pattern,
|
1058 |
+
image_size=image_size,
|
1059 |
+
allow_empty_phrase=True,
|
1060 |
+
)
|
1061 |
+
parsed_dict['polygons'] = instances
|
1062 |
+
elif task == 'bboxes':
|
1063 |
+
instances = self.parse_description_with_bboxes_from_text_and_spans(
|
1064 |
+
text,
|
1065 |
+
pattern=pattern,
|
1066 |
+
image_size=image_size,
|
1067 |
+
allow_empty_phrase=True,
|
1068 |
+
)
|
1069 |
+
parsed_dict['bboxes'] = instances
|
1070 |
+
elif task == 'description_with_bboxes_or_polygons':
|
1071 |
+
if '<poly>' in text:
|
1072 |
+
# only support either polygons or bboxes, not both at the same time
|
1073 |
+
instances = self.parse_description_with_polygons_from_text_and_spans(
|
1074 |
+
text,
|
1075 |
+
pattern=pattern,
|
1076 |
+
image_size=image_size,
|
1077 |
+
)
|
1078 |
+
else:
|
1079 |
+
instances = self.parse_description_with_bboxes_from_text_and_spans(
|
1080 |
+
text,
|
1081 |
+
pattern=pattern,
|
1082 |
+
image_size=image_size,
|
1083 |
+
)
|
1084 |
+
parsed_dict['description_with_bboxes_or_polygons'] = instances
|
1085 |
+
else:
|
1086 |
+
raise ValueError("task {} is not supported".format(task))
|
1087 |
+
|
1088 |
+
return parsed_dict
|
processor_config.json
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoProcessor": "processing_florence2.Florence2Processor"
|
4 |
+
},
|
5 |
+
"processor_class": "Florence2Processor"
|
6 |
+
}
|
special_tokens_map.json
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": {
|
3 |
+
"content": "<bos>",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": false,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"eos_token": {
|
10 |
+
"content": "<eos>",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": false,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"pad_token": {
|
17 |
+
"content": "<pad>",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": false,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
},
|
23 |
+
"unk_token": {
|
24 |
+
"content": "<unk>",
|
25 |
+
"lstrip": false,
|
26 |
+
"normalized": false,
|
27 |
+
"rstrip": false,
|
28 |
+
"single_word": false
|
29 |
+
}
|
30 |
+
}
|
tokenizer.json
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3f289bc05132635a8bc7aca7aa21255efd5e18f3710f43e3cdb96bcd41be4922
|
3 |
+
size 17525357
|
tokenizer_config.json
ADDED
@@ -0,0 +1,2010 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"add_bos_token": true,
|
3 |
+
"add_eos_token": false,
|
4 |
+
"added_tokens_decoder": {
|
5 |
+
"0": {
|
6 |
+
"content": "<pad>",
|
7 |
+
"lstrip": false,
|
8 |
+
"normalized": false,
|
9 |
+
"rstrip": false,
|
10 |
+
"single_word": false,
|
11 |
+
"special": true
|
12 |
+
},
|
13 |
+
"1": {
|
14 |
+
"content": "<eos>",
|
15 |
+
"lstrip": false,
|
16 |
+
"normalized": false,
|
17 |
+
"rstrip": false,
|
18 |
+
"single_word": false,
|
19 |
+
"special": true
|
20 |
+
},
|
21 |
+
"2": {
|
22 |
+
"content": "<bos>",
|
23 |
+
"lstrip": false,
|
24 |
+
"normalized": false,
|
25 |
+
"rstrip": false,
|
26 |
+
"single_word": false,
|
27 |
+
"special": true
|
28 |
+
},
|
29 |
+
"3": {
|
30 |
+
"content": "<unk>",
|
31 |
+
"lstrip": false,
|
32 |
+
"normalized": false,
|
33 |
+
"rstrip": false,
|
34 |
+
"single_word": false,
|
35 |
+
"special": true
|
36 |
+
},
|
37 |
+
"4": {
|
38 |
+
"content": "<mask>",
|
39 |
+
"lstrip": false,
|
40 |
+
"normalized": false,
|
41 |
+
"rstrip": false,
|
42 |
+
"single_word": false,
|
43 |
+
"special": false
|
44 |
+
},
|
45 |
+
"5": {
|
46 |
+
"content": "<2mass>",
|
47 |
+
"lstrip": false,
|
48 |
+
"normalized": false,
|
49 |
+
"rstrip": false,
|
50 |
+
"single_word": false,
|
51 |
+
"special": false
|
52 |
+
},
|
53 |
+
"6": {
|
54 |
+
"content": "[@BOS@]",
|
55 |
+
"lstrip": false,
|
56 |
+
"normalized": false,
|
57 |
+
"rstrip": false,
|
58 |
+
"single_word": false,
|
59 |
+
"special": false
|
60 |
+
},
|
61 |
+
"7": {
|
62 |
+
"content": "<unused0>",
|
63 |
+
"lstrip": false,
|
64 |
+
"normalized": false,
|
65 |
+
"rstrip": false,
|
66 |
+
"single_word": false,
|
67 |
+
"special": false
|
68 |
+
},
|
69 |
+
"8": {
|
70 |
+
"content": "<unused1>",
|
71 |
+
"lstrip": false,
|
72 |
+
"normalized": false,
|
73 |
+
"rstrip": false,
|
74 |
+
"single_word": false,
|
75 |
+
"special": false
|
76 |
+
},
|
77 |
+
"9": {
|
78 |
+
"content": "<unused2>",
|
79 |
+
"lstrip": false,
|
80 |
+
"normalized": false,
|
81 |
+
"rstrip": false,
|
82 |
+
"single_word": false,
|
83 |
+
"special": false
|
84 |
+
},
|
85 |
+
"10": {
|
86 |
+
"content": "<unused3>",
|
87 |
+
"lstrip": false,
|
88 |
+
"normalized": false,
|
89 |
+
"rstrip": false,
|
90 |
+
"single_word": false,
|
91 |
+
"special": false
|
92 |
+
},
|
93 |
+
"11": {
|
94 |
+
"content": "<unused4>",
|
95 |
+
"lstrip": false,
|
96 |
+
"normalized": false,
|
97 |
+
"rstrip": false,
|
98 |
+
"single_word": false,
|
99 |
+
"special": false
|
100 |
+
},
|
101 |
+
"12": {
|
102 |
+
"content": "<unused5>",
|
103 |
+
"lstrip": false,
|
104 |
+
"normalized": false,
|
105 |
+
"rstrip": false,
|
106 |
+
"single_word": false,
|
107 |
+
"special": false
|
108 |
+
},
|
109 |
+
"13": {
|
110 |
+
"content": "<unused6>",
|
111 |
+
"lstrip": false,
|
112 |
+
"normalized": false,
|
113 |
+
"rstrip": false,
|
114 |
+
"single_word": false,
|
115 |
+
"special": false
|
116 |
+
},
|
117 |
+
"14": {
|
118 |
+
"content": "<unused7>",
|
119 |
+
"lstrip": false,
|
120 |
+
"normalized": false,
|
121 |
+
"rstrip": false,
|
122 |
+
"single_word": false,
|
123 |
+
"special": false
|
124 |
+
},
|
125 |
+
"15": {
|
126 |
+
"content": "<unused8>",
|
127 |
+
"lstrip": false,
|
128 |
+
"normalized": false,
|
129 |
+
"rstrip": false,
|
130 |
+
"single_word": false,
|
131 |
+
"special": false
|
132 |
+
},
|
133 |
+
"16": {
|
134 |
+
"content": "<unused9>",
|
135 |
+
"lstrip": false,
|
136 |
+
"normalized": false,
|
137 |
+
"rstrip": false,
|
138 |
+
"single_word": false,
|
139 |
+
"special": false
|
140 |
+
},
|
141 |
+
"17": {
|
142 |
+
"content": "<unused10>",
|
143 |
+
"lstrip": false,
|
144 |
+
"normalized": false,
|
145 |
+
"rstrip": false,
|
146 |
+
"single_word": false,
|
147 |
+
"special": false
|
148 |
+
},
|
149 |
+
"18": {
|
150 |
+
"content": "<unused11>",
|
151 |
+
"lstrip": false,
|
152 |
+
"normalized": false,
|
153 |
+
"rstrip": false,
|
154 |
+
"single_word": false,
|
155 |
+
"special": false
|
156 |
+
},
|
157 |
+
"19": {
|
158 |
+
"content": "<unused12>",
|
159 |
+
"lstrip": false,
|
160 |
+
"normalized": false,
|
161 |
+
"rstrip": false,
|
162 |
+
"single_word": false,
|
163 |
+
"special": false
|
164 |
+
},
|
165 |
+
"20": {
|
166 |
+
"content": "<unused13>",
|
167 |
+
"lstrip": false,
|
168 |
+
"normalized": false,
|
169 |
+
"rstrip": false,
|
170 |
+
"single_word": false,
|
171 |
+
"special": false
|
172 |
+
},
|
173 |
+
"21": {
|
174 |
+
"content": "<unused14>",
|
175 |
+
"lstrip": false,
|
176 |
+
"normalized": false,
|
177 |
+
"rstrip": false,
|
178 |
+
"single_word": false,
|
179 |
+
"special": false
|
180 |
+
},
|
181 |
+
"22": {
|
182 |
+
"content": "<unused15>",
|
183 |
+
"lstrip": false,
|
184 |
+
"normalized": false,
|
185 |
+
"rstrip": false,
|
186 |
+
"single_word": false,
|
187 |
+
"special": false
|
188 |
+
},
|
189 |
+
"23": {
|
190 |
+
"content": "<unused16>",
|
191 |
+
"lstrip": false,
|
192 |
+
"normalized": false,
|
193 |
+
"rstrip": false,
|
194 |
+
"single_word": false,
|
195 |
+
"special": false
|
196 |
+
},
|
197 |
+
"24": {
|
198 |
+
"content": "<unused17>",
|
199 |
+
"lstrip": false,
|
200 |
+
"normalized": false,
|
201 |
+
"rstrip": false,
|
202 |
+
"single_word": false,
|
203 |
+
"special": false
|
204 |
+
},
|
205 |
+
"25": {
|
206 |
+
"content": "<unused18>",
|
207 |
+
"lstrip": false,
|
208 |
+
"normalized": false,
|
209 |
+
"rstrip": false,
|
210 |
+
"single_word": false,
|
211 |
+
"special": false
|
212 |
+
},
|
213 |
+
"26": {
|
214 |
+
"content": "<unused19>",
|
215 |
+
"lstrip": false,
|
216 |
+
"normalized": false,
|
217 |
+
"rstrip": false,
|
218 |
+
"single_word": false,
|
219 |
+
"special": false
|
220 |
+
},
|
221 |
+
"27": {
|
222 |
+
"content": "<unused20>",
|
223 |
+
"lstrip": false,
|
224 |
+
"normalized": false,
|
225 |
+
"rstrip": false,
|
226 |
+
"single_word": false,
|
227 |
+
"special": false
|
228 |
+
},
|
229 |
+
"28": {
|
230 |
+
"content": "<unused21>",
|
231 |
+
"lstrip": false,
|
232 |
+
"normalized": false,
|
233 |
+
"rstrip": false,
|
234 |
+
"single_word": false,
|
235 |
+
"special": false
|
236 |
+
},
|
237 |
+
"29": {
|
238 |
+
"content": "<unused22>",
|
239 |
+
"lstrip": false,
|
240 |
+
"normalized": false,
|
241 |
+
"rstrip": false,
|
242 |
+
"single_word": false,
|
243 |
+
"special": false
|
244 |
+
},
|
245 |
+
"30": {
|
246 |
+
"content": "<unused23>",
|
247 |
+
"lstrip": false,
|
248 |
+
"normalized": false,
|
249 |
+
"rstrip": false,
|
250 |
+
"single_word": false,
|
251 |
+
"special": false
|
252 |
+
},
|
253 |
+
"31": {
|
254 |
+
"content": "<unused24>",
|
255 |
+
"lstrip": false,
|
256 |
+
"normalized": false,
|
257 |
+
"rstrip": false,
|
258 |
+
"single_word": false,
|
259 |
+
"special": false
|
260 |
+
},
|
261 |
+
"32": {
|
262 |
+
"content": "<unused25>",
|
263 |
+
"lstrip": false,
|
264 |
+
"normalized": false,
|
265 |
+
"rstrip": false,
|
266 |
+
"single_word": false,
|
267 |
+
"special": false
|
268 |
+
},
|
269 |
+
"33": {
|
270 |
+
"content": "<unused26>",
|
271 |
+
"lstrip": false,
|
272 |
+
"normalized": false,
|
273 |
+
"rstrip": false,
|
274 |
+
"single_word": false,
|
275 |
+
"special": false
|
276 |
+
},
|
277 |
+
"34": {
|
278 |
+
"content": "<unused27>",
|
279 |
+
"lstrip": false,
|
280 |
+
"normalized": false,
|
281 |
+
"rstrip": false,
|
282 |
+
"single_word": false,
|
283 |
+
"special": false
|
284 |
+
},
|
285 |
+
"35": {
|
286 |
+
"content": "<unused28>",
|
287 |
+
"lstrip": false,
|
288 |
+
"normalized": false,
|
289 |
+
"rstrip": false,
|
290 |
+
"single_word": false,
|
291 |
+
"special": false
|
292 |
+
},
|
293 |
+
"36": {
|
294 |
+
"content": "<unused29>",
|
295 |
+
"lstrip": false,
|
296 |
+
"normalized": false,
|
297 |
+
"rstrip": false,
|
298 |
+
"single_word": false,
|
299 |
+
"special": false
|
300 |
+
},
|
301 |
+
"37": {
|
302 |
+
"content": "<unused30>",
|
303 |
+
"lstrip": false,
|
304 |
+
"normalized": false,
|
305 |
+
"rstrip": false,
|
306 |
+
"single_word": false,
|
307 |
+
"special": false
|
308 |
+
},
|
309 |
+
"38": {
|
310 |
+
"content": "<unused31>",
|
311 |
+
"lstrip": false,
|
312 |
+
"normalized": false,
|
313 |
+
"rstrip": false,
|
314 |
+
"single_word": false,
|
315 |
+
"special": false
|
316 |
+
},
|
317 |
+
"39": {
|
318 |
+
"content": "<unused32>",
|
319 |
+
"lstrip": false,
|
320 |
+
"normalized": false,
|
321 |
+
"rstrip": false,
|
322 |
+
"single_word": false,
|
323 |
+
"special": false
|
324 |
+
},
|
325 |
+
"40": {
|
326 |
+
"content": "<unused33>",
|
327 |
+
"lstrip": false,
|
328 |
+
"normalized": false,
|
329 |
+
"rstrip": false,
|
330 |
+
"single_word": false,
|
331 |
+
"special": false
|
332 |
+
},
|
333 |
+
"41": {
|
334 |
+
"content": "<unused34>",
|
335 |
+
"lstrip": false,
|
336 |
+
"normalized": false,
|
337 |
+
"rstrip": false,
|
338 |
+
"single_word": false,
|
339 |
+
"special": false
|
340 |
+
},
|
341 |
+
"42": {
|
342 |
+
"content": "<unused35>",
|
343 |
+
"lstrip": false,
|
344 |
+
"normalized": false,
|
345 |
+
"rstrip": false,
|
346 |
+
"single_word": false,
|
347 |
+
"special": false
|
348 |
+
},
|
349 |
+
"43": {
|
350 |
+
"content": "<unused36>",
|
351 |
+
"lstrip": false,
|
352 |
+
"normalized": false,
|
353 |
+
"rstrip": false,
|
354 |
+
"single_word": false,
|
355 |
+
"special": false
|
356 |
+
},
|
357 |
+
"44": {
|
358 |
+
"content": "<unused37>",
|
359 |
+
"lstrip": false,
|
360 |
+
"normalized": false,
|
361 |
+
"rstrip": false,
|
362 |
+
"single_word": false,
|
363 |
+
"special": false
|
364 |
+
},
|
365 |
+
"45": {
|
366 |
+
"content": "<unused38>",
|
367 |
+
"lstrip": false,
|
368 |
+
"normalized": false,
|
369 |
+
"rstrip": false,
|
370 |
+
"single_word": false,
|
371 |
+
"special": false
|
372 |
+
},
|
373 |
+
"46": {
|
374 |
+
"content": "<unused39>",
|
375 |
+
"lstrip": false,
|
376 |
+
"normalized": false,
|
377 |
+
"rstrip": false,
|
378 |
+
"single_word": false,
|
379 |
+
"special": false
|
380 |
+
},
|
381 |
+
"47": {
|
382 |
+
"content": "<unused40>",
|
383 |
+
"lstrip": false,
|
384 |
+
"normalized": false,
|
385 |
+
"rstrip": false,
|
386 |
+
"single_word": false,
|
387 |
+
"special": false
|
388 |
+
},
|
389 |
+
"48": {
|
390 |
+
"content": "<unused41>",
|
391 |
+
"lstrip": false,
|
392 |
+
"normalized": false,
|
393 |
+
"rstrip": false,
|
394 |
+
"single_word": false,
|
395 |
+
"special": false
|
396 |
+
},
|
397 |
+
"49": {
|
398 |
+
"content": "<unused42>",
|
399 |
+
"lstrip": false,
|
400 |
+
"normalized": false,
|
401 |
+
"rstrip": false,
|
402 |
+
"single_word": false,
|
403 |
+
"special": false
|
404 |
+
},
|
405 |
+
"50": {
|
406 |
+
"content": "<unused43>",
|
407 |
+
"lstrip": false,
|
408 |
+
"normalized": false,
|
409 |
+
"rstrip": false,
|
410 |
+
"single_word": false,
|
411 |
+
"special": false
|
412 |
+
},
|
413 |
+
"51": {
|
414 |
+
"content": "<unused44>",
|
415 |
+
"lstrip": false,
|
416 |
+
"normalized": false,
|
417 |
+
"rstrip": false,
|
418 |
+
"single_word": false,
|
419 |
+
"special": false
|
420 |
+
},
|
421 |
+
"52": {
|
422 |
+
"content": "<unused45>",
|
423 |
+
"lstrip": false,
|
424 |
+
"normalized": false,
|
425 |
+
"rstrip": false,
|
426 |
+
"single_word": false,
|
427 |
+
"special": false
|
428 |
+
},
|
429 |
+
"53": {
|
430 |
+
"content": "<unused46>",
|
431 |
+
"lstrip": false,
|
432 |
+
"normalized": false,
|
433 |
+
"rstrip": false,
|
434 |
+
"single_word": false,
|
435 |
+
"special": false
|
436 |
+
},
|
437 |
+
"54": {
|
438 |
+
"content": "<unused47>",
|
439 |
+
"lstrip": false,
|
440 |
+
"normalized": false,
|
441 |
+
"rstrip": false,
|
442 |
+
"single_word": false,
|
443 |
+
"special": false
|
444 |
+
},
|
445 |
+
"55": {
|
446 |
+
"content": "<unused48>",
|
447 |
+
"lstrip": false,
|
448 |
+
"normalized": false,
|
449 |
+
"rstrip": false,
|
450 |
+
"single_word": false,
|
451 |
+
"special": false
|
452 |
+
},
|
453 |
+
"56": {
|
454 |
+
"content": "<unused49>",
|
455 |
+
"lstrip": false,
|
456 |
+
"normalized": false,
|
457 |
+
"rstrip": false,
|
458 |
+
"single_word": false,
|
459 |
+
"special": false
|
460 |
+
},
|
461 |
+
"57": {
|
462 |
+
"content": "<unused50>",
|
463 |
+
"lstrip": false,
|
464 |
+
"normalized": false,
|
465 |
+
"rstrip": false,
|
466 |
+
"single_word": false,
|
467 |
+
"special": false
|
468 |
+
},
|
469 |
+
"58": {
|
470 |
+
"content": "<unused51>",
|
471 |
+
"lstrip": false,
|
472 |
+
"normalized": false,
|
473 |
+
"rstrip": false,
|
474 |
+
"single_word": false,
|
475 |
+
"special": false
|
476 |
+
},
|
477 |
+
"59": {
|
478 |
+
"content": "<unused52>",
|
479 |
+
"lstrip": false,
|
480 |
+
"normalized": false,
|
481 |
+
"rstrip": false,
|
482 |
+
"single_word": false,
|
483 |
+
"special": false
|
484 |
+
},
|
485 |
+
"60": {
|
486 |
+
"content": "<unused53>",
|
487 |
+
"lstrip": false,
|
488 |
+
"normalized": false,
|
489 |
+
"rstrip": false,
|
490 |
+
"single_word": false,
|
491 |
+
"special": false
|
492 |
+
},
|
493 |
+
"61": {
|
494 |
+
"content": "<unused54>",
|
495 |
+
"lstrip": false,
|
496 |
+
"normalized": false,
|
497 |
+
"rstrip": false,
|
498 |
+
"single_word": false,
|
499 |
+
"special": false
|
500 |
+
},
|
501 |
+
"62": {
|
502 |
+
"content": "<unused55>",
|
503 |
+
"lstrip": false,
|
504 |
+
"normalized": false,
|
505 |
+
"rstrip": false,
|
506 |
+
"single_word": false,
|
507 |
+
"special": false
|
508 |
+
},
|
509 |
+
"63": {
|
510 |
+
"content": "<unused56>",
|
511 |
+
"lstrip": false,
|
512 |
+
"normalized": false,
|
513 |
+
"rstrip": false,
|
514 |
+
"single_word": false,
|
515 |
+
"special": false
|
516 |
+
},
|
517 |
+
"64": {
|
518 |
+
"content": "<unused57>",
|
519 |
+
"lstrip": false,
|
520 |
+
"normalized": false,
|
521 |
+
"rstrip": false,
|
522 |
+
"single_word": false,
|
523 |
+
"special": false
|
524 |
+
},
|
525 |
+
"65": {
|
526 |
+
"content": "<unused58>",
|
527 |
+
"lstrip": false,
|
528 |
+
"normalized": false,
|
529 |
+
"rstrip": false,
|
530 |
+
"single_word": false,
|
531 |
+
"special": false
|
532 |
+
},
|
533 |
+
"66": {
|
534 |
+
"content": "<unused59>",
|
535 |
+
"lstrip": false,
|
536 |
+
"normalized": false,
|
537 |
+
"rstrip": false,
|
538 |
+
"single_word": false,
|
539 |
+
"special": false
|
540 |
+
},
|
541 |
+
"67": {
|
542 |
+
"content": "<unused60>",
|
543 |
+
"lstrip": false,
|
544 |
+
"normalized": false,
|
545 |
+
"rstrip": false,
|
546 |
+
"single_word": false,
|
547 |
+
"special": false
|
548 |
+
},
|
549 |
+
"68": {
|
550 |
+
"content": "<unused61>",
|
551 |
+
"lstrip": false,
|
552 |
+
"normalized": false,
|
553 |
+
"rstrip": false,
|
554 |
+
"single_word": false,
|
555 |
+
"special": false
|
556 |
+
},
|
557 |
+
"69": {
|
558 |
+
"content": "<unused62>",
|
559 |
+
"lstrip": false,
|
560 |
+
"normalized": false,
|
561 |
+
"rstrip": false,
|
562 |
+
"single_word": false,
|
563 |
+
"special": false
|
564 |
+
},
|
565 |
+
"70": {
|
566 |
+
"content": "<unused63>",
|
567 |
+
"lstrip": false,
|
568 |
+
"normalized": false,
|
569 |
+
"rstrip": false,
|
570 |
+
"single_word": false,
|
571 |
+
"special": false
|
572 |
+
},
|
573 |
+
"71": {
|
574 |
+
"content": "<unused64>",
|
575 |
+
"lstrip": false,
|
576 |
+
"normalized": false,
|
577 |
+
"rstrip": false,
|
578 |
+
"single_word": false,
|
579 |
+
"special": false
|
580 |
+
},
|
581 |
+
"72": {
|
582 |
+
"content": "<unused65>",
|
583 |
+
"lstrip": false,
|
584 |
+
"normalized": false,
|
585 |
+
"rstrip": false,
|
586 |
+
"single_word": false,
|
587 |
+
"special": false
|
588 |
+
},
|
589 |
+
"73": {
|
590 |
+
"content": "<unused66>",
|
591 |
+
"lstrip": false,
|
592 |
+
"normalized": false,
|
593 |
+
"rstrip": false,
|
594 |
+
"single_word": false,
|
595 |
+
"special": false
|
596 |
+
},
|
597 |
+
"74": {
|
598 |
+
"content": "<unused67>",
|
599 |
+
"lstrip": false,
|
600 |
+
"normalized": false,
|
601 |
+
"rstrip": false,
|
602 |
+
"single_word": false,
|
603 |
+
"special": false
|
604 |
+
},
|
605 |
+
"75": {
|
606 |
+
"content": "<unused68>",
|
607 |
+
"lstrip": false,
|
608 |
+
"normalized": false,
|
609 |
+
"rstrip": false,
|
610 |
+
"single_word": false,
|
611 |
+
"special": false
|
612 |
+
},
|
613 |
+
"76": {
|
614 |
+
"content": "<unused69>",
|
615 |
+
"lstrip": false,
|
616 |
+
"normalized": false,
|
617 |
+
"rstrip": false,
|
618 |
+
"single_word": false,
|
619 |
+
"special": false
|
620 |
+
},
|
621 |
+
"77": {
|
622 |
+
"content": "<unused70>",
|
623 |
+
"lstrip": false,
|
624 |
+
"normalized": false,
|
625 |
+
"rstrip": false,
|
626 |
+
"single_word": false,
|
627 |
+
"special": false
|
628 |
+
},
|
629 |
+
"78": {
|
630 |
+
"content": "<unused71>",
|
631 |
+
"lstrip": false,
|
632 |
+
"normalized": false,
|
633 |
+
"rstrip": false,
|
634 |
+
"single_word": false,
|
635 |
+
"special": false
|
636 |
+
},
|
637 |
+
"79": {
|
638 |
+
"content": "<unused72>",
|
639 |
+
"lstrip": false,
|
640 |
+
"normalized": false,
|
641 |
+
"rstrip": false,
|
642 |
+
"single_word": false,
|
643 |
+
"special": false
|
644 |
+
},
|
645 |
+
"80": {
|
646 |
+
"content": "<unused73>",
|
647 |
+
"lstrip": false,
|
648 |
+
"normalized": false,
|
649 |
+
"rstrip": false,
|
650 |
+
"single_word": false,
|
651 |
+
"special": false
|
652 |
+
},
|
653 |
+
"81": {
|
654 |
+
"content": "<unused74>",
|
655 |
+
"lstrip": false,
|
656 |
+
"normalized": false,
|
657 |
+
"rstrip": false,
|
658 |
+
"single_word": false,
|
659 |
+
"special": false
|
660 |
+
},
|
661 |
+
"82": {
|
662 |
+
"content": "<unused75>",
|
663 |
+
"lstrip": false,
|
664 |
+
"normalized": false,
|
665 |
+
"rstrip": false,
|
666 |
+
"single_word": false,
|
667 |
+
"special": false
|
668 |
+
},
|
669 |
+
"83": {
|
670 |
+
"content": "<unused76>",
|
671 |
+
"lstrip": false,
|
672 |
+
"normalized": false,
|
673 |
+
"rstrip": false,
|
674 |
+
"single_word": false,
|
675 |
+
"special": false
|
676 |
+
},
|
677 |
+
"84": {
|
678 |
+
"content": "<unused77>",
|
679 |
+
"lstrip": false,
|
680 |
+
"normalized": false,
|
681 |
+
"rstrip": false,
|
682 |
+
"single_word": false,
|
683 |
+
"special": false
|
684 |
+
},
|
685 |
+
"85": {
|
686 |
+
"content": "<unused78>",
|
687 |
+
"lstrip": false,
|
688 |
+
"normalized": false,
|
689 |
+
"rstrip": false,
|
690 |
+
"single_word": false,
|
691 |
+
"special": false
|
692 |
+
},
|
693 |
+
"86": {
|
694 |
+
"content": "<unused79>",
|
695 |
+
"lstrip": false,
|
696 |
+
"normalized": false,
|
697 |
+
"rstrip": false,
|
698 |
+
"single_word": false,
|
699 |
+
"special": false
|
700 |
+
},
|
701 |
+
"87": {
|
702 |
+
"content": "<unused80>",
|
703 |
+
"lstrip": false,
|
704 |
+
"normalized": false,
|
705 |
+
"rstrip": false,
|
706 |
+
"single_word": false,
|
707 |
+
"special": false
|
708 |
+
},
|
709 |
+
"88": {
|
710 |
+
"content": "<unused81>",
|
711 |
+
"lstrip": false,
|
712 |
+
"normalized": false,
|
713 |
+
"rstrip": false,
|
714 |
+
"single_word": false,
|
715 |
+
"special": false
|
716 |
+
},
|
717 |
+
"89": {
|
718 |
+
"content": "<unused82>",
|
719 |
+
"lstrip": false,
|
720 |
+
"normalized": false,
|
721 |
+
"rstrip": false,
|
722 |
+
"single_word": false,
|
723 |
+
"special": false
|
724 |
+
},
|
725 |
+
"90": {
|
726 |
+
"content": "<unused83>",
|
727 |
+
"lstrip": false,
|
728 |
+
"normalized": false,
|
729 |
+
"rstrip": false,
|
730 |
+
"single_word": false,
|
731 |
+
"special": false
|
732 |
+
},
|
733 |
+
"91": {
|
734 |
+
"content": "<unused84>",
|
735 |
+
"lstrip": false,
|
736 |
+
"normalized": false,
|
737 |
+
"rstrip": false,
|
738 |
+
"single_word": false,
|
739 |
+
"special": false
|
740 |
+
},
|
741 |
+
"92": {
|
742 |
+
"content": "<unused85>",
|
743 |
+
"lstrip": false,
|
744 |
+
"normalized": false,
|
745 |
+
"rstrip": false,
|
746 |
+
"single_word": false,
|
747 |
+
"special": false
|
748 |
+
},
|
749 |
+
"93": {
|
750 |
+
"content": "<unused86>",
|
751 |
+
"lstrip": false,
|
752 |
+
"normalized": false,
|
753 |
+
"rstrip": false,
|
754 |
+
"single_word": false,
|
755 |
+
"special": false
|
756 |
+
},
|
757 |
+
"94": {
|
758 |
+
"content": "<unused87>",
|
759 |
+
"lstrip": false,
|
760 |
+
"normalized": false,
|
761 |
+
"rstrip": false,
|
762 |
+
"single_word": false,
|
763 |
+
"special": false
|
764 |
+
},
|
765 |
+
"95": {
|
766 |
+
"content": "<unused88>",
|
767 |
+
"lstrip": false,
|
768 |
+
"normalized": false,
|
769 |
+
"rstrip": false,
|
770 |
+
"single_word": false,
|
771 |
+
"special": false
|
772 |
+
},
|
773 |
+
"96": {
|
774 |
+
"content": "<unused89>",
|
775 |
+
"lstrip": false,
|
776 |
+
"normalized": false,
|
777 |
+
"rstrip": false,
|
778 |
+
"single_word": false,
|
779 |
+
"special": false
|
780 |
+
},
|
781 |
+
"97": {
|
782 |
+
"content": "<unused90>",
|
783 |
+
"lstrip": false,
|
784 |
+
"normalized": false,
|
785 |
+
"rstrip": false,
|
786 |
+
"single_word": false,
|
787 |
+
"special": false
|
788 |
+
},
|
789 |
+
"98": {
|
790 |
+
"content": "<unused91>",
|
791 |
+
"lstrip": false,
|
792 |
+
"normalized": false,
|
793 |
+
"rstrip": false,
|
794 |
+
"single_word": false,
|
795 |
+
"special": false
|
796 |
+
},
|
797 |
+
"99": {
|
798 |
+
"content": "<unused92>",
|
799 |
+
"lstrip": false,
|
800 |
+
"normalized": false,
|
801 |
+
"rstrip": false,
|
802 |
+
"single_word": false,
|
803 |
+
"special": false
|
804 |
+
},
|
805 |
+
"100": {
|
806 |
+
"content": "<unused93>",
|
807 |
+
"lstrip": false,
|
808 |
+
"normalized": false,
|
809 |
+
"rstrip": false,
|
810 |
+
"single_word": false,
|
811 |
+
"special": false
|
812 |
+
},
|
813 |
+
"101": {
|
814 |
+
"content": "<unused94>",
|
815 |
+
"lstrip": false,
|
816 |
+
"normalized": false,
|
817 |
+
"rstrip": false,
|
818 |
+
"single_word": false,
|
819 |
+
"special": false
|
820 |
+
},
|
821 |
+
"102": {
|
822 |
+
"content": "<unused95>",
|
823 |
+
"lstrip": false,
|
824 |
+
"normalized": false,
|
825 |
+
"rstrip": false,
|
826 |
+
"single_word": false,
|
827 |
+
"special": false
|
828 |
+
},
|
829 |
+
"103": {
|
830 |
+
"content": "<unused96>",
|
831 |
+
"lstrip": false,
|
832 |
+
"normalized": false,
|
833 |
+
"rstrip": false,
|
834 |
+
"single_word": false,
|
835 |
+
"special": false
|
836 |
+
},
|
837 |
+
"104": {
|
838 |
+
"content": "<unused97>",
|
839 |
+
"lstrip": false,
|
840 |
+
"normalized": false,
|
841 |
+
"rstrip": false,
|
842 |
+
"single_word": false,
|
843 |
+
"special": false
|
844 |
+
},
|
845 |
+
"105": {
|
846 |
+
"content": "<unused98>",
|
847 |
+
"lstrip": false,
|
848 |
+
"normalized": false,
|
849 |
+
"rstrip": false,
|
850 |
+
"single_word": false,
|
851 |
+
"special": false
|
852 |
+
},
|
853 |
+
"106": {
|
854 |
+
"content": "<start_of_turn>",
|
855 |
+
"lstrip": false,
|
856 |
+
"normalized": false,
|
857 |
+
"rstrip": false,
|
858 |
+
"single_word": false,
|
859 |
+
"special": true
|
860 |
+
},
|
861 |
+
"107": {
|
862 |
+
"content": "<end_of_turn>",
|
863 |
+
"lstrip": false,
|
864 |
+
"normalized": false,
|
865 |
+
"rstrip": false,
|
866 |
+
"single_word": false,
|
867 |
+
"special": true
|
868 |
+
},
|
869 |
+
"108": {
|
870 |
+
"content": "\n",
|
871 |
+
"lstrip": false,
|
872 |
+
"normalized": false,
|
873 |
+
"rstrip": false,
|
874 |
+
"single_word": false,
|
875 |
+
"special": false
|
876 |
+
},
|
877 |
+
"109": {
|
878 |
+
"content": "\n\n",
|
879 |
+
"lstrip": false,
|
880 |
+
"normalized": false,
|
881 |
+
"rstrip": false,
|
882 |
+
"single_word": false,
|
883 |
+
"special": false
|
884 |
+
},
|
885 |
+
"110": {
|
886 |
+
"content": "\n\n\n",
|
887 |
+
"lstrip": false,
|
888 |
+
"normalized": false,
|
889 |
+
"rstrip": false,
|
890 |
+
"single_word": false,
|
891 |
+
"special": false
|
892 |
+
},
|
893 |
+
"111": {
|
894 |
+
"content": "\n\n\n\n",
|
895 |
+
"lstrip": false,
|
896 |
+
"normalized": false,
|
897 |
+
"rstrip": false,
|
898 |
+
"single_word": false,
|
899 |
+
"special": false
|
900 |
+
},
|
901 |
+
"112": {
|
902 |
+
"content": "\n\n\n\n\n",
|
903 |
+
"lstrip": false,
|
904 |
+
"normalized": false,
|
905 |
+
"rstrip": false,
|
906 |
+
"single_word": false,
|
907 |
+
"special": false
|
908 |
+
},
|
909 |
+
"113": {
|
910 |
+
"content": "\n\n\n\n\n\n",
|
911 |
+
"lstrip": false,
|
912 |
+
"normalized": false,
|
913 |
+
"rstrip": false,
|
914 |
+
"single_word": false,
|
915 |
+
"special": false
|
916 |
+
},
|
917 |
+
"114": {
|
918 |
+
"content": "\n\n\n\n\n\n\n",
|
919 |
+
"lstrip": false,
|
920 |
+
"normalized": false,
|
921 |
+
"rstrip": false,
|
922 |
+
"single_word": false,
|
923 |
+
"special": false
|
924 |
+
},
|
925 |
+
"115": {
|
926 |
+
"content": "\n\n\n\n\n\n\n\n",
|
927 |
+
"lstrip": false,
|
928 |
+
"normalized": false,
|
929 |
+
"rstrip": false,
|
930 |
+
"single_word": false,
|
931 |
+
"special": false
|
932 |
+
},
|
933 |
+
"116": {
|
934 |
+
"content": "\n\n\n\n\n\n\n\n\n",
|
935 |
+
"lstrip": false,
|
936 |
+
"normalized": false,
|
937 |
+
"rstrip": false,
|
938 |
+
"single_word": false,
|
939 |
+
"special": false
|
940 |
+
},
|
941 |
+
"117": {
|
942 |
+
"content": "\n\n\n\n\n\n\n\n\n\n",
|
943 |
+
"lstrip": false,
|
944 |
+
"normalized": false,
|
945 |
+
"rstrip": false,
|
946 |
+
"single_word": false,
|
947 |
+
"special": false
|
948 |
+
},
|
949 |
+
"118": {
|
950 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n",
|
951 |
+
"lstrip": false,
|
952 |
+
"normalized": false,
|
953 |
+
"rstrip": false,
|
954 |
+
"single_word": false,
|
955 |
+
"special": false
|
956 |
+
},
|
957 |
+
"119": {
|
958 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n",
|
959 |
+
"lstrip": false,
|
960 |
+
"normalized": false,
|
961 |
+
"rstrip": false,
|
962 |
+
"single_word": false,
|
963 |
+
"special": false
|
964 |
+
},
|
965 |
+
"120": {
|
966 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
967 |
+
"lstrip": false,
|
968 |
+
"normalized": false,
|
969 |
+
"rstrip": false,
|
970 |
+
"single_word": false,
|
971 |
+
"special": false
|
972 |
+
},
|
973 |
+
"121": {
|
974 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
975 |
+
"lstrip": false,
|
976 |
+
"normalized": false,
|
977 |
+
"rstrip": false,
|
978 |
+
"single_word": false,
|
979 |
+
"special": false
|
980 |
+
},
|
981 |
+
"122": {
|
982 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
983 |
+
"lstrip": false,
|
984 |
+
"normalized": false,
|
985 |
+
"rstrip": false,
|
986 |
+
"single_word": false,
|
987 |
+
"special": false
|
988 |
+
},
|
989 |
+
"123": {
|
990 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
991 |
+
"lstrip": false,
|
992 |
+
"normalized": false,
|
993 |
+
"rstrip": false,
|
994 |
+
"single_word": false,
|
995 |
+
"special": false
|
996 |
+
},
|
997 |
+
"124": {
|
998 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
999 |
+
"lstrip": false,
|
1000 |
+
"normalized": false,
|
1001 |
+
"rstrip": false,
|
1002 |
+
"single_word": false,
|
1003 |
+
"special": false
|
1004 |
+
},
|
1005 |
+
"125": {
|
1006 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1007 |
+
"lstrip": false,
|
1008 |
+
"normalized": false,
|
1009 |
+
"rstrip": false,
|
1010 |
+
"single_word": false,
|
1011 |
+
"special": false
|
1012 |
+
},
|
1013 |
+
"126": {
|
1014 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1015 |
+
"lstrip": false,
|
1016 |
+
"normalized": false,
|
1017 |
+
"rstrip": false,
|
1018 |
+
"single_word": false,
|
1019 |
+
"special": false
|
1020 |
+
},
|
1021 |
+
"127": {
|
1022 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1023 |
+
"lstrip": false,
|
1024 |
+
"normalized": false,
|
1025 |
+
"rstrip": false,
|
1026 |
+
"single_word": false,
|
1027 |
+
"special": false
|
1028 |
+
},
|
1029 |
+
"128": {
|
1030 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1031 |
+
"lstrip": false,
|
1032 |
+
"normalized": false,
|
1033 |
+
"rstrip": false,
|
1034 |
+
"single_word": false,
|
1035 |
+
"special": false
|
1036 |
+
},
|
1037 |
+
"129": {
|
1038 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1039 |
+
"lstrip": false,
|
1040 |
+
"normalized": false,
|
1041 |
+
"rstrip": false,
|
1042 |
+
"single_word": false,
|
1043 |
+
"special": false
|
1044 |
+
},
|
1045 |
+
"130": {
|
1046 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1047 |
+
"lstrip": false,
|
1048 |
+
"normalized": false,
|
1049 |
+
"rstrip": false,
|
1050 |
+
"single_word": false,
|
1051 |
+
"special": false
|
1052 |
+
},
|
1053 |
+
"131": {
|
1054 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1055 |
+
"lstrip": false,
|
1056 |
+
"normalized": false,
|
1057 |
+
"rstrip": false,
|
1058 |
+
"single_word": false,
|
1059 |
+
"special": false
|
1060 |
+
},
|
1061 |
+
"132": {
|
1062 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1063 |
+
"lstrip": false,
|
1064 |
+
"normalized": false,
|
1065 |
+
"rstrip": false,
|
1066 |
+
"single_word": false,
|
1067 |
+
"special": false
|
1068 |
+
},
|
1069 |
+
"133": {
|
1070 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1071 |
+
"lstrip": false,
|
1072 |
+
"normalized": false,
|
1073 |
+
"rstrip": false,
|
1074 |
+
"single_word": false,
|
1075 |
+
"special": false
|
1076 |
+
},
|
1077 |
+
"134": {
|
1078 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1079 |
+
"lstrip": false,
|
1080 |
+
"normalized": false,
|
1081 |
+
"rstrip": false,
|
1082 |
+
"single_word": false,
|
1083 |
+
"special": false
|
1084 |
+
},
|
1085 |
+
"135": {
|
1086 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1087 |
+
"lstrip": false,
|
1088 |
+
"normalized": false,
|
1089 |
+
"rstrip": false,
|
1090 |
+
"single_word": false,
|
1091 |
+
"special": false
|
1092 |
+
},
|
1093 |
+
"136": {
|
1094 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1095 |
+
"lstrip": false,
|
1096 |
+
"normalized": false,
|
1097 |
+
"rstrip": false,
|
1098 |
+
"single_word": false,
|
1099 |
+
"special": false
|
1100 |
+
},
|
1101 |
+
"137": {
|
1102 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1103 |
+
"lstrip": false,
|
1104 |
+
"normalized": false,
|
1105 |
+
"rstrip": false,
|
1106 |
+
"single_word": false,
|
1107 |
+
"special": false
|
1108 |
+
},
|
1109 |
+
"138": {
|
1110 |
+
"content": "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
|
1111 |
+
"lstrip": false,
|
1112 |
+
"normalized": false,
|
1113 |
+
"rstrip": false,
|
1114 |
+
"single_word": false,
|
1115 |
+
"special": false
|
1116 |
+
},
|
1117 |
+
"139": {
|
1118 |
+
"content": "▁▁",
|
1119 |
+
"lstrip": false,
|
1120 |
+
"normalized": false,
|
1121 |
+
"rstrip": false,
|
1122 |
+
"single_word": false,
|
1123 |
+
"special": false
|
1124 |
+
},
|
1125 |
+
"140": {
|
1126 |
+
"content": "▁▁▁",
|
1127 |
+
"lstrip": false,
|
1128 |
+
"normalized": false,
|
1129 |
+
"rstrip": false,
|
1130 |
+
"single_word": false,
|
1131 |
+
"special": false
|
1132 |
+
},
|
1133 |
+
"141": {
|
1134 |
+
"content": "▁▁▁▁",
|
1135 |
+
"lstrip": false,
|
1136 |
+
"normalized": false,
|
1137 |
+
"rstrip": false,
|
1138 |
+
"single_word": false,
|
1139 |
+
"special": false
|
1140 |
+
},
|
1141 |
+
"142": {
|
1142 |
+
"content": "▁▁▁▁▁",
|
1143 |
+
"lstrip": false,
|
1144 |
+
"normalized": false,
|
1145 |
+
"rstrip": false,
|
1146 |
+
"single_word": false,
|
1147 |
+
"special": false
|
1148 |
+
},
|
1149 |
+
"143": {
|
1150 |
+
"content": "▁▁▁▁▁▁",
|
1151 |
+
"lstrip": false,
|
1152 |
+
"normalized": false,
|
1153 |
+
"rstrip": false,
|
1154 |
+
"single_word": false,
|
1155 |
+
"special": false
|
1156 |
+
},
|
1157 |
+
"144": {
|
1158 |
+
"content": "▁▁▁▁▁▁▁",
|
1159 |
+
"lstrip": false,
|
1160 |
+
"normalized": false,
|
1161 |
+
"rstrip": false,
|
1162 |
+
"single_word": false,
|
1163 |
+
"special": false
|
1164 |
+
},
|
1165 |
+
"145": {
|
1166 |
+
"content": "▁▁▁▁▁▁▁▁",
|
1167 |
+
"lstrip": false,
|
1168 |
+
"normalized": false,
|
1169 |
+
"rstrip": false,
|
1170 |
+
"single_word": false,
|
1171 |
+
"special": false
|
1172 |
+
},
|
1173 |
+
"146": {
|
1174 |
+
"content": "▁▁▁▁▁▁▁▁▁",
|
1175 |
+
"lstrip": false,
|
1176 |
+
"normalized": false,
|
1177 |
+
"rstrip": false,
|
1178 |
+
"single_word": false,
|
1179 |
+
"special": false
|
1180 |
+
},
|
1181 |
+
"147": {
|
1182 |
+
"content": "▁▁▁▁▁▁▁▁▁▁",
|
1183 |
+
"lstrip": false,
|
1184 |
+
"normalized": false,
|
1185 |
+
"rstrip": false,
|
1186 |
+
"single_word": false,
|
1187 |
+
"special": false
|
1188 |
+
},
|
1189 |
+
"148": {
|
1190 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁",
|
1191 |
+
"lstrip": false,
|
1192 |
+
"normalized": false,
|
1193 |
+
"rstrip": false,
|
1194 |
+
"single_word": false,
|
1195 |
+
"special": false
|
1196 |
+
},
|
1197 |
+
"149": {
|
1198 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁",
|
1199 |
+
"lstrip": false,
|
1200 |
+
"normalized": false,
|
1201 |
+
"rstrip": false,
|
1202 |
+
"single_word": false,
|
1203 |
+
"special": false
|
1204 |
+
},
|
1205 |
+
"150": {
|
1206 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1207 |
+
"lstrip": false,
|
1208 |
+
"normalized": false,
|
1209 |
+
"rstrip": false,
|
1210 |
+
"single_word": false,
|
1211 |
+
"special": false
|
1212 |
+
},
|
1213 |
+
"151": {
|
1214 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1215 |
+
"lstrip": false,
|
1216 |
+
"normalized": false,
|
1217 |
+
"rstrip": false,
|
1218 |
+
"single_word": false,
|
1219 |
+
"special": false
|
1220 |
+
},
|
1221 |
+
"152": {
|
1222 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1223 |
+
"lstrip": false,
|
1224 |
+
"normalized": false,
|
1225 |
+
"rstrip": false,
|
1226 |
+
"single_word": false,
|
1227 |
+
"special": false
|
1228 |
+
},
|
1229 |
+
"153": {
|
1230 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1231 |
+
"lstrip": false,
|
1232 |
+
"normalized": false,
|
1233 |
+
"rstrip": false,
|
1234 |
+
"single_word": false,
|
1235 |
+
"special": false
|
1236 |
+
},
|
1237 |
+
"154": {
|
1238 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1239 |
+
"lstrip": false,
|
1240 |
+
"normalized": false,
|
1241 |
+
"rstrip": false,
|
1242 |
+
"single_word": false,
|
1243 |
+
"special": false
|
1244 |
+
},
|
1245 |
+
"155": {
|
1246 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1247 |
+
"lstrip": false,
|
1248 |
+
"normalized": false,
|
1249 |
+
"rstrip": false,
|
1250 |
+
"single_word": false,
|
1251 |
+
"special": false
|
1252 |
+
},
|
1253 |
+
"156": {
|
1254 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1255 |
+
"lstrip": false,
|
1256 |
+
"normalized": false,
|
1257 |
+
"rstrip": false,
|
1258 |
+
"single_word": false,
|
1259 |
+
"special": false
|
1260 |
+
},
|
1261 |
+
"157": {
|
1262 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1263 |
+
"lstrip": false,
|
1264 |
+
"normalized": false,
|
1265 |
+
"rstrip": false,
|
1266 |
+
"single_word": false,
|
1267 |
+
"special": false
|
1268 |
+
},
|
1269 |
+
"158": {
|
1270 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1271 |
+
"lstrip": false,
|
1272 |
+
"normalized": false,
|
1273 |
+
"rstrip": false,
|
1274 |
+
"single_word": false,
|
1275 |
+
"special": false
|
1276 |
+
},
|
1277 |
+
"159": {
|
1278 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1279 |
+
"lstrip": false,
|
1280 |
+
"normalized": false,
|
1281 |
+
"rstrip": false,
|
1282 |
+
"single_word": false,
|
1283 |
+
"special": false
|
1284 |
+
},
|
1285 |
+
"160": {
|
1286 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1287 |
+
"lstrip": false,
|
1288 |
+
"normalized": false,
|
1289 |
+
"rstrip": false,
|
1290 |
+
"single_word": false,
|
1291 |
+
"special": false
|
1292 |
+
},
|
1293 |
+
"161": {
|
1294 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1295 |
+
"lstrip": false,
|
1296 |
+
"normalized": false,
|
1297 |
+
"rstrip": false,
|
1298 |
+
"single_word": false,
|
1299 |
+
"special": false
|
1300 |
+
},
|
1301 |
+
"162": {
|
1302 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1303 |
+
"lstrip": false,
|
1304 |
+
"normalized": false,
|
1305 |
+
"rstrip": false,
|
1306 |
+
"single_word": false,
|
1307 |
+
"special": false
|
1308 |
+
},
|
1309 |
+
"163": {
|
1310 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1311 |
+
"lstrip": false,
|
1312 |
+
"normalized": false,
|
1313 |
+
"rstrip": false,
|
1314 |
+
"single_word": false,
|
1315 |
+
"special": false
|
1316 |
+
},
|
1317 |
+
"164": {
|
1318 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1319 |
+
"lstrip": false,
|
1320 |
+
"normalized": false,
|
1321 |
+
"rstrip": false,
|
1322 |
+
"single_word": false,
|
1323 |
+
"special": false
|
1324 |
+
},
|
1325 |
+
"165": {
|
1326 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1327 |
+
"lstrip": false,
|
1328 |
+
"normalized": false,
|
1329 |
+
"rstrip": false,
|
1330 |
+
"single_word": false,
|
1331 |
+
"special": false
|
1332 |
+
},
|
1333 |
+
"166": {
|
1334 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1335 |
+
"lstrip": false,
|
1336 |
+
"normalized": false,
|
1337 |
+
"rstrip": false,
|
1338 |
+
"single_word": false,
|
1339 |
+
"special": false
|
1340 |
+
},
|
1341 |
+
"167": {
|
1342 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1343 |
+
"lstrip": false,
|
1344 |
+
"normalized": false,
|
1345 |
+
"rstrip": false,
|
1346 |
+
"single_word": false,
|
1347 |
+
"special": false
|
1348 |
+
},
|
1349 |
+
"168": {
|
1350 |
+
"content": "▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁",
|
1351 |
+
"lstrip": false,
|
1352 |
+
"normalized": false,
|
1353 |
+
"rstrip": false,
|
1354 |
+
"single_word": false,
|
1355 |
+
"special": false
|
1356 |
+
},
|
1357 |
+
"169": {
|
1358 |
+
"content": "<table>",
|
1359 |
+
"lstrip": false,
|
1360 |
+
"normalized": false,
|
1361 |
+
"rstrip": false,
|
1362 |
+
"single_word": false,
|
1363 |
+
"special": false
|
1364 |
+
},
|
1365 |
+
"170": {
|
1366 |
+
"content": "<caption>",
|
1367 |
+
"lstrip": false,
|
1368 |
+
"normalized": false,
|
1369 |
+
"rstrip": false,
|
1370 |
+
"single_word": false,
|
1371 |
+
"special": false
|
1372 |
+
},
|
1373 |
+
"171": {
|
1374 |
+
"content": "<thead>",
|
1375 |
+
"lstrip": false,
|
1376 |
+
"normalized": false,
|
1377 |
+
"rstrip": false,
|
1378 |
+
"single_word": false,
|
1379 |
+
"special": false
|
1380 |
+
},
|
1381 |
+
"172": {
|
1382 |
+
"content": "<tbody>",
|
1383 |
+
"lstrip": false,
|
1384 |
+
"normalized": false,
|
1385 |
+
"rstrip": false,
|
1386 |
+
"single_word": false,
|
1387 |
+
"special": false
|
1388 |
+
},
|
1389 |
+
"173": {
|
1390 |
+
"content": "<tfoot>",
|
1391 |
+
"lstrip": false,
|
1392 |
+
"normalized": false,
|
1393 |
+
"rstrip": false,
|
1394 |
+
"single_word": false,
|
1395 |
+
"special": false
|
1396 |
+
},
|
1397 |
+
"174": {
|
1398 |
+
"content": "<tr>",
|
1399 |
+
"lstrip": false,
|
1400 |
+
"normalized": false,
|
1401 |
+
"rstrip": false,
|
1402 |
+
"single_word": false,
|
1403 |
+
"special": false
|
1404 |
+
},
|
1405 |
+
"175": {
|
1406 |
+
"content": "<th>",
|
1407 |
+
"lstrip": false,
|
1408 |
+
"normalized": false,
|
1409 |
+
"rstrip": false,
|
1410 |
+
"single_word": false,
|
1411 |
+
"special": false
|
1412 |
+
},
|
1413 |
+
"176": {
|
1414 |
+
"content": "<td>",
|
1415 |
+
"lstrip": false,
|
1416 |
+
"normalized": false,
|
1417 |
+
"rstrip": false,
|
1418 |
+
"single_word": false,
|
1419 |
+
"special": false
|
1420 |
+
},
|
1421 |
+
"177": {
|
1422 |
+
"content": "</table>",
|
1423 |
+
"lstrip": false,
|
1424 |
+
"normalized": false,
|
1425 |
+
"rstrip": false,
|
1426 |
+
"single_word": false,
|
1427 |
+
"special": false
|
1428 |
+
},
|
1429 |
+
"178": {
|
1430 |
+
"content": "</caption>",
|
1431 |
+
"lstrip": false,
|
1432 |
+
"normalized": false,
|
1433 |
+
"rstrip": false,
|
1434 |
+
"single_word": false,
|
1435 |
+
"special": false
|
1436 |
+
},
|
1437 |
+
"179": {
|
1438 |
+
"content": "</thead>",
|
1439 |
+
"lstrip": false,
|
1440 |
+
"normalized": false,
|
1441 |
+
"rstrip": false,
|
1442 |
+
"single_word": false,
|
1443 |
+
"special": false
|
1444 |
+
},
|
1445 |
+
"180": {
|
1446 |
+
"content": "</tbody>",
|
1447 |
+
"lstrip": false,
|
1448 |
+
"normalized": false,
|
1449 |
+
"rstrip": false,
|
1450 |
+
"single_word": false,
|
1451 |
+
"special": false
|
1452 |
+
},
|
1453 |
+
"181": {
|
1454 |
+
"content": "</tfoot>",
|
1455 |
+
"lstrip": false,
|
1456 |
+
"normalized": false,
|
1457 |
+
"rstrip": false,
|
1458 |
+
"single_word": false,
|
1459 |
+
"special": false
|
1460 |
+
},
|
1461 |
+
"182": {
|
1462 |
+
"content": "</tr>",
|
1463 |
+
"lstrip": false,
|
1464 |
+
"normalized": false,
|
1465 |
+
"rstrip": false,
|
1466 |
+
"single_word": false,
|
1467 |
+
"special": false
|
1468 |
+
},
|
1469 |
+
"183": {
|
1470 |
+
"content": "</th>",
|
1471 |
+
"lstrip": false,
|
1472 |
+
"normalized": false,
|
1473 |
+
"rstrip": false,
|
1474 |
+
"single_word": false,
|
1475 |
+
"special": false
|
1476 |
+
},
|
1477 |
+
"184": {
|
1478 |
+
"content": "</td>",
|
1479 |
+
"lstrip": false,
|
1480 |
+
"normalized": false,
|
1481 |
+
"rstrip": false,
|
1482 |
+
"single_word": false,
|
1483 |
+
"special": false
|
1484 |
+
},
|
1485 |
+
"185": {
|
1486 |
+
"content": "<h1>",
|
1487 |
+
"lstrip": false,
|
1488 |
+
"normalized": false,
|
1489 |
+
"rstrip": false,
|
1490 |
+
"single_word": false,
|
1491 |
+
"special": false
|
1492 |
+
},
|
1493 |
+
"186": {
|
1494 |
+
"content": "<h2>",
|
1495 |
+
"lstrip": false,
|
1496 |
+
"normalized": false,
|
1497 |
+
"rstrip": false,
|
1498 |
+
"single_word": false,
|
1499 |
+
"special": false
|
1500 |
+
},
|
1501 |
+
"187": {
|
1502 |
+
"content": "<h3>",
|
1503 |
+
"lstrip": false,
|
1504 |
+
"normalized": false,
|
1505 |
+
"rstrip": false,
|
1506 |
+
"single_word": false,
|
1507 |
+
"special": false
|
1508 |
+
},
|
1509 |
+
"188": {
|
1510 |
+
"content": "<h4>",
|
1511 |
+
"lstrip": false,
|
1512 |
+
"normalized": false,
|
1513 |
+
"rstrip": false,
|
1514 |
+
"single_word": false,
|
1515 |
+
"special": false
|
1516 |
+
},
|
1517 |
+
"189": {
|
1518 |
+
"content": "<h5>",
|
1519 |
+
"lstrip": false,
|
1520 |
+
"normalized": false,
|
1521 |
+
"rstrip": false,
|
1522 |
+
"single_word": false,
|
1523 |
+
"special": false
|
1524 |
+
},
|
1525 |
+
"190": {
|
1526 |
+
"content": "<h6>",
|
1527 |
+
"lstrip": false,
|
1528 |
+
"normalized": false,
|
1529 |
+
"rstrip": false,
|
1530 |
+
"single_word": false,
|
1531 |
+
"special": false
|
1532 |
+
},
|
1533 |
+
"191": {
|
1534 |
+
"content": "<blockquote>",
|
1535 |
+
"lstrip": false,
|
1536 |
+
"normalized": false,
|
1537 |
+
"rstrip": false,
|
1538 |
+
"single_word": false,
|
1539 |
+
"special": false
|
1540 |
+
},
|
1541 |
+
"192": {
|
1542 |
+
"content": "</h1>",
|
1543 |
+
"lstrip": false,
|
1544 |
+
"normalized": false,
|
1545 |
+
"rstrip": false,
|
1546 |
+
"single_word": false,
|
1547 |
+
"special": false
|
1548 |
+
},
|
1549 |
+
"193": {
|
1550 |
+
"content": "</h2>",
|
1551 |
+
"lstrip": false,
|
1552 |
+
"normalized": false,
|
1553 |
+
"rstrip": false,
|
1554 |
+
"single_word": false,
|
1555 |
+
"special": false
|
1556 |
+
},
|
1557 |
+
"194": {
|
1558 |
+
"content": "</h3>",
|
1559 |
+
"lstrip": false,
|
1560 |
+
"normalized": false,
|
1561 |
+
"rstrip": false,
|
1562 |
+
"single_word": false,
|
1563 |
+
"special": false
|
1564 |
+
},
|
1565 |
+
"195": {
|
1566 |
+
"content": "</h4>",
|
1567 |
+
"lstrip": false,
|
1568 |
+
"normalized": false,
|
1569 |
+
"rstrip": false,
|
1570 |
+
"single_word": false,
|
1571 |
+
"special": false
|
1572 |
+
},
|
1573 |
+
"196": {
|
1574 |
+
"content": "</h5>",
|
1575 |
+
"lstrip": false,
|
1576 |
+
"normalized": false,
|
1577 |
+
"rstrip": false,
|
1578 |
+
"single_word": false,
|
1579 |
+
"special": false
|
1580 |
+
},
|
1581 |
+
"197": {
|
1582 |
+
"content": "</h6>",
|
1583 |
+
"lstrip": false,
|
1584 |
+
"normalized": false,
|
1585 |
+
"rstrip": false,
|
1586 |
+
"single_word": false,
|
1587 |
+
"special": false
|
1588 |
+
},
|
1589 |
+
"198": {
|
1590 |
+
"content": "</blockquote>",
|
1591 |
+
"lstrip": false,
|
1592 |
+
"normalized": false,
|
1593 |
+
"rstrip": false,
|
1594 |
+
"single_word": false,
|
1595 |
+
"special": false
|
1596 |
+
},
|
1597 |
+
"199": {
|
1598 |
+
"content": "<strong>",
|
1599 |
+
"lstrip": false,
|
1600 |
+
"normalized": false,
|
1601 |
+
"rstrip": false,
|
1602 |
+
"single_word": false,
|
1603 |
+
"special": false
|
1604 |
+
},
|
1605 |
+
"200": {
|
1606 |
+
"content": "<em>",
|
1607 |
+
"lstrip": false,
|
1608 |
+
"normalized": false,
|
1609 |
+
"rstrip": false,
|
1610 |
+
"single_word": false,
|
1611 |
+
"special": false
|
1612 |
+
},
|
1613 |
+
"201": {
|
1614 |
+
"content": "<b>",
|
1615 |
+
"lstrip": false,
|
1616 |
+
"normalized": false,
|
1617 |
+
"rstrip": false,
|
1618 |
+
"single_word": false,
|
1619 |
+
"special": false
|
1620 |
+
},
|
1621 |
+
"202": {
|
1622 |
+
"content": "<i>",
|
1623 |
+
"lstrip": false,
|
1624 |
+
"normalized": false,
|
1625 |
+
"rstrip": false,
|
1626 |
+
"single_word": false,
|
1627 |
+
"special": false
|
1628 |
+
},
|
1629 |
+
"203": {
|
1630 |
+
"content": "<u>",
|
1631 |
+
"lstrip": false,
|
1632 |
+
"normalized": false,
|
1633 |
+
"rstrip": false,
|
1634 |
+
"single_word": false,
|
1635 |
+
"special": false
|
1636 |
+
},
|
1637 |
+
"204": {
|
1638 |
+
"content": "<s>",
|
1639 |
+
"lstrip": false,
|
1640 |
+
"normalized": false,
|
1641 |
+
"rstrip": false,
|
1642 |
+
"single_word": false,
|
1643 |
+
"special": false
|
1644 |
+
},
|
1645 |
+
"205": {
|
1646 |
+
"content": "<sub>",
|
1647 |
+
"lstrip": false,
|
1648 |
+
"normalized": false,
|
1649 |
+
"rstrip": false,
|
1650 |
+
"single_word": false,
|
1651 |
+
"special": false
|
1652 |
+
},
|
1653 |
+
"206": {
|
1654 |
+
"content": "<sup>",
|
1655 |
+
"lstrip": false,
|
1656 |
+
"normalized": false,
|
1657 |
+
"rstrip": false,
|
1658 |
+
"single_word": false,
|
1659 |
+
"special": false
|
1660 |
+
},
|
1661 |
+
"207": {
|
1662 |
+
"content": "<code>",
|
1663 |
+
"lstrip": false,
|
1664 |
+
"normalized": false,
|
1665 |
+
"rstrip": false,
|
1666 |
+
"single_word": false,
|
1667 |
+
"special": false
|
1668 |
+
},
|
1669 |
+
"208": {
|
1670 |
+
"content": "</strong>",
|
1671 |
+
"lstrip": false,
|
1672 |
+
"normalized": false,
|
1673 |
+
"rstrip": false,
|
1674 |
+
"single_word": false,
|
1675 |
+
"special": false
|
1676 |
+
},
|
1677 |
+
"209": {
|
1678 |
+
"content": "</em>",
|
1679 |
+
"lstrip": false,
|
1680 |
+
"normalized": false,
|
1681 |
+
"rstrip": false,
|
1682 |
+
"single_word": false,
|
1683 |
+
"special": false
|
1684 |
+
},
|
1685 |
+
"210": {
|
1686 |
+
"content": "</b>",
|
1687 |
+
"lstrip": false,
|
1688 |
+
"normalized": false,
|
1689 |
+
"rstrip": false,
|
1690 |
+
"single_word": false,
|
1691 |
+
"special": false
|
1692 |
+
},
|
1693 |
+
"211": {
|
1694 |
+
"content": "</i>",
|
1695 |
+
"lstrip": false,
|
1696 |
+
"normalized": false,
|
1697 |
+
"rstrip": false,
|
1698 |
+
"single_word": false,
|
1699 |
+
"special": false
|
1700 |
+
},
|
1701 |
+
"212": {
|
1702 |
+
"content": "</u>",
|
1703 |
+
"lstrip": false,
|
1704 |
+
"normalized": false,
|
1705 |
+
"rstrip": false,
|
1706 |
+
"single_word": false,
|
1707 |
+
"special": false
|
1708 |
+
},
|
1709 |
+
"213": {
|
1710 |
+
"content": "</s>",
|
1711 |
+
"lstrip": false,
|
1712 |
+
"normalized": false,
|
1713 |
+
"rstrip": false,
|
1714 |
+
"single_word": false,
|
1715 |
+
"special": false
|
1716 |
+
},
|
1717 |
+
"214": {
|
1718 |
+
"content": "</sub>",
|
1719 |
+
"lstrip": false,
|
1720 |
+
"normalized": false,
|
1721 |
+
"rstrip": false,
|
1722 |
+
"single_word": false,
|
1723 |
+
"special": false
|
1724 |
+
},
|
1725 |
+
"215": {
|
1726 |
+
"content": "</sup>",
|
1727 |
+
"lstrip": false,
|
1728 |
+
"normalized": false,
|
1729 |
+
"rstrip": false,
|
1730 |
+
"single_word": false,
|
1731 |
+
"special": false
|
1732 |
+
},
|
1733 |
+
"216": {
|
1734 |
+
"content": "</code>",
|
1735 |
+
"lstrip": false,
|
1736 |
+
"normalized": false,
|
1737 |
+
"rstrip": false,
|
1738 |
+
"single_word": false,
|
1739 |
+
"special": false
|
1740 |
+
},
|
1741 |
+
"255968": {
|
1742 |
+
"content": "[toxicity=0]",
|
1743 |
+
"lstrip": false,
|
1744 |
+
"normalized": false,
|
1745 |
+
"rstrip": false,
|
1746 |
+
"single_word": false,
|
1747 |
+
"special": false
|
1748 |
+
},
|
1749 |
+
"255969": {
|
1750 |
+
"content": "\t\t",
|
1751 |
+
"lstrip": false,
|
1752 |
+
"normalized": false,
|
1753 |
+
"rstrip": false,
|
1754 |
+
"single_word": false,
|
1755 |
+
"special": false
|
1756 |
+
},
|
1757 |
+
"255970": {
|
1758 |
+
"content": "\t\t\t",
|
1759 |
+
"lstrip": false,
|
1760 |
+
"normalized": false,
|
1761 |
+
"rstrip": false,
|
1762 |
+
"single_word": false,
|
1763 |
+
"special": false
|
1764 |
+
},
|
1765 |
+
"255971": {
|
1766 |
+
"content": "\t\t\t\t",
|
1767 |
+
"lstrip": false,
|
1768 |
+
"normalized": false,
|
1769 |
+
"rstrip": false,
|
1770 |
+
"single_word": false,
|
1771 |
+
"special": false
|
1772 |
+
},
|
1773 |
+
"255972": {
|
1774 |
+
"content": "\t\t\t\t\t",
|
1775 |
+
"lstrip": false,
|
1776 |
+
"normalized": false,
|
1777 |
+
"rstrip": false,
|
1778 |
+
"single_word": false,
|
1779 |
+
"special": false
|
1780 |
+
},
|
1781 |
+
"255973": {
|
1782 |
+
"content": "\t\t\t\t\t\t",
|
1783 |
+
"lstrip": false,
|
1784 |
+
"normalized": false,
|
1785 |
+
"rstrip": false,
|
1786 |
+
"single_word": false,
|
1787 |
+
"special": false
|
1788 |
+
},
|
1789 |
+
"255974": {
|
1790 |
+
"content": "\t\t\t\t\t\t\t",
|
1791 |
+
"lstrip": false,
|
1792 |
+
"normalized": false,
|
1793 |
+
"rstrip": false,
|
1794 |
+
"single_word": false,
|
1795 |
+
"special": false
|
1796 |
+
},
|
1797 |
+
"255975": {
|
1798 |
+
"content": "\t\t\t\t\t\t\t\t",
|
1799 |
+
"lstrip": false,
|
1800 |
+
"normalized": false,
|
1801 |
+
"rstrip": false,
|
1802 |
+
"single_word": false,
|
1803 |
+
"special": false
|
1804 |
+
},
|
1805 |
+
"255976": {
|
1806 |
+
"content": "\t\t\t\t\t\t\t\t\t",
|
1807 |
+
"lstrip": false,
|
1808 |
+
"normalized": false,
|
1809 |
+
"rstrip": false,
|
1810 |
+
"single_word": false,
|
1811 |
+
"special": false
|
1812 |
+
},
|
1813 |
+
"255977": {
|
1814 |
+
"content": "\t\t\t\t\t\t\t\t\t\t",
|
1815 |
+
"lstrip": false,
|
1816 |
+
"normalized": false,
|
1817 |
+
"rstrip": false,
|
1818 |
+
"single_word": false,
|
1819 |
+
"special": false
|
1820 |
+
},
|
1821 |
+
"255978": {
|
1822 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t",
|
1823 |
+
"lstrip": false,
|
1824 |
+
"normalized": false,
|
1825 |
+
"rstrip": false,
|
1826 |
+
"single_word": false,
|
1827 |
+
"special": false
|
1828 |
+
},
|
1829 |
+
"255979": {
|
1830 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t",
|
1831 |
+
"lstrip": false,
|
1832 |
+
"normalized": false,
|
1833 |
+
"rstrip": false,
|
1834 |
+
"single_word": false,
|
1835 |
+
"special": false
|
1836 |
+
},
|
1837 |
+
"255980": {
|
1838 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1839 |
+
"lstrip": false,
|
1840 |
+
"normalized": false,
|
1841 |
+
"rstrip": false,
|
1842 |
+
"single_word": false,
|
1843 |
+
"special": false
|
1844 |
+
},
|
1845 |
+
"255981": {
|
1846 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1847 |
+
"lstrip": false,
|
1848 |
+
"normalized": false,
|
1849 |
+
"rstrip": false,
|
1850 |
+
"single_word": false,
|
1851 |
+
"special": false
|
1852 |
+
},
|
1853 |
+
"255982": {
|
1854 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1855 |
+
"lstrip": false,
|
1856 |
+
"normalized": false,
|
1857 |
+
"rstrip": false,
|
1858 |
+
"single_word": false,
|
1859 |
+
"special": false
|
1860 |
+
},
|
1861 |
+
"255983": {
|
1862 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1863 |
+
"lstrip": false,
|
1864 |
+
"normalized": false,
|
1865 |
+
"rstrip": false,
|
1866 |
+
"single_word": false,
|
1867 |
+
"special": false
|
1868 |
+
},
|
1869 |
+
"255984": {
|
1870 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1871 |
+
"lstrip": false,
|
1872 |
+
"normalized": false,
|
1873 |
+
"rstrip": false,
|
1874 |
+
"single_word": false,
|
1875 |
+
"special": false
|
1876 |
+
},
|
1877 |
+
"255985": {
|
1878 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1879 |
+
"lstrip": false,
|
1880 |
+
"normalized": false,
|
1881 |
+
"rstrip": false,
|
1882 |
+
"single_word": false,
|
1883 |
+
"special": false
|
1884 |
+
},
|
1885 |
+
"255986": {
|
1886 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1887 |
+
"lstrip": false,
|
1888 |
+
"normalized": false,
|
1889 |
+
"rstrip": false,
|
1890 |
+
"single_word": false,
|
1891 |
+
"special": false
|
1892 |
+
},
|
1893 |
+
"255987": {
|
1894 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1895 |
+
"lstrip": false,
|
1896 |
+
"normalized": false,
|
1897 |
+
"rstrip": false,
|
1898 |
+
"single_word": false,
|
1899 |
+
"special": false
|
1900 |
+
},
|
1901 |
+
"255988": {
|
1902 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1903 |
+
"lstrip": false,
|
1904 |
+
"normalized": false,
|
1905 |
+
"rstrip": false,
|
1906 |
+
"single_word": false,
|
1907 |
+
"special": false
|
1908 |
+
},
|
1909 |
+
"255989": {
|
1910 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1911 |
+
"lstrip": false,
|
1912 |
+
"normalized": false,
|
1913 |
+
"rstrip": false,
|
1914 |
+
"single_word": false,
|
1915 |
+
"special": false
|
1916 |
+
},
|
1917 |
+
"255990": {
|
1918 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1919 |
+
"lstrip": false,
|
1920 |
+
"normalized": false,
|
1921 |
+
"rstrip": false,
|
1922 |
+
"single_word": false,
|
1923 |
+
"special": false
|
1924 |
+
},
|
1925 |
+
"255991": {
|
1926 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1927 |
+
"lstrip": false,
|
1928 |
+
"normalized": false,
|
1929 |
+
"rstrip": false,
|
1930 |
+
"single_word": false,
|
1931 |
+
"special": false
|
1932 |
+
},
|
1933 |
+
"255992": {
|
1934 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1935 |
+
"lstrip": false,
|
1936 |
+
"normalized": false,
|
1937 |
+
"rstrip": false,
|
1938 |
+
"single_word": false,
|
1939 |
+
"special": false
|
1940 |
+
},
|
1941 |
+
"255993": {
|
1942 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1943 |
+
"lstrip": false,
|
1944 |
+
"normalized": false,
|
1945 |
+
"rstrip": false,
|
1946 |
+
"single_word": false,
|
1947 |
+
"special": false
|
1948 |
+
},
|
1949 |
+
"255994": {
|
1950 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1951 |
+
"lstrip": false,
|
1952 |
+
"normalized": false,
|
1953 |
+
"rstrip": false,
|
1954 |
+
"single_word": false,
|
1955 |
+
"special": false
|
1956 |
+
},
|
1957 |
+
"255995": {
|
1958 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1959 |
+
"lstrip": false,
|
1960 |
+
"normalized": false,
|
1961 |
+
"rstrip": false,
|
1962 |
+
"single_word": false,
|
1963 |
+
"special": false
|
1964 |
+
},
|
1965 |
+
"255996": {
|
1966 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1967 |
+
"lstrip": false,
|
1968 |
+
"normalized": false,
|
1969 |
+
"rstrip": false,
|
1970 |
+
"single_word": false,
|
1971 |
+
"special": false
|
1972 |
+
},
|
1973 |
+
"255997": {
|
1974 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1975 |
+
"lstrip": false,
|
1976 |
+
"normalized": false,
|
1977 |
+
"rstrip": false,
|
1978 |
+
"single_word": false,
|
1979 |
+
"special": false
|
1980 |
+
},
|
1981 |
+
"255998": {
|
1982 |
+
"content": "\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t",
|
1983 |
+
"lstrip": false,
|
1984 |
+
"normalized": false,
|
1985 |
+
"rstrip": false,
|
1986 |
+
"single_word": false,
|
1987 |
+
"special": false
|
1988 |
+
},
|
1989 |
+
"255999": {
|
1990 |
+
"content": "<unused99>",
|
1991 |
+
"lstrip": false,
|
1992 |
+
"normalized": false,
|
1993 |
+
"rstrip": false,
|
1994 |
+
"single_word": false,
|
1995 |
+
"special": false
|
1996 |
+
}
|
1997 |
+
},
|
1998 |
+
"bos_token": "<bos>",
|
1999 |
+
"chat_template": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}",
|
2000 |
+
"clean_up_tokenization_spaces": false,
|
2001 |
+
"eos_token": "<eos>",
|
2002 |
+
"model_max_length": 1000000000000000019884624838656,
|
2003 |
+
"pad_token": "<pad>",
|
2004 |
+
"processor_class": "Florence2Processor",
|
2005 |
+
"sp_model_kwargs": {},
|
2006 |
+
"spaces_between_special_tokens": false,
|
2007 |
+
"tokenizer_class": "GemmaTokenizer",
|
2008 |
+
"unk_token": "<unk>",
|
2009 |
+
"use_default_system_prompt": false
|
2010 |
+
}
|