update
Browse files- modeling_qwen.py +10 -153
- tokenization_qwen.py +27 -7
- visual.py +70 -19
modeling_qwen.py
CHANGED
@@ -69,44 +69,7 @@ Pass argument `stream` to model.chat() is buggy, deprecated, and marked for remo
|
|
69 |
|
70 |
apply_rotary_emb_func = None
|
71 |
rms_norm = None
|
72 |
-
flash_attn_unpadded_func = None
|
73 |
-
|
74 |
-
|
75 |
-
def _import_flash_attn():
|
76 |
-
global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
|
77 |
-
try:
|
78 |
-
from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
|
79 |
-
apply_rotary_emb_func = __apply_rotary_emb_func
|
80 |
-
except ImportError:
|
81 |
-
logger.warn(
|
82 |
-
"Warning: import flash_attn rotary fail, please install FlashAttention rotary to get higher efficiency "
|
83 |
-
"https://github.com/Dao-AILab/flash-attention/tree/main/csrc/rotary"
|
84 |
-
)
|
85 |
|
86 |
-
try:
|
87 |
-
from flash_attn.ops.rms_norm import rms_norm as __rms_norm
|
88 |
-
rms_norm = __rms_norm
|
89 |
-
except ImportError:
|
90 |
-
logger.warn(
|
91 |
-
"Warning: import flash_attn rms_norm fail, please install FlashAttention layer_norm to get higher efficiency "
|
92 |
-
"https://github.com/Dao-AILab/flash-attention/tree/main/csrc/layer_norm"
|
93 |
-
)
|
94 |
-
|
95 |
-
try:
|
96 |
-
import flash_attn
|
97 |
-
if not hasattr(flash_attn, '__version__'):
|
98 |
-
from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
|
99 |
-
else:
|
100 |
-
if int(flash_attn.__version__.split(".")[0]) >= 2:
|
101 |
-
from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
|
102 |
-
else:
|
103 |
-
from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
|
104 |
-
flash_attn_unpadded_func = __flash_attn_unpadded_func
|
105 |
-
except ImportError:
|
106 |
-
logger.warn(
|
107 |
-
"Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
|
108 |
-
"https://github.com/Dao-AILab/flash-attention"
|
109 |
-
)
|
110 |
|
111 |
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
112 |
def _make_causal_mask(
|
@@ -141,70 +104,6 @@ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int]
|
|
141 |
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
142 |
|
143 |
|
144 |
-
class FlashSelfAttention(torch.nn.Module):
|
145 |
-
def __init__(
|
146 |
-
self,
|
147 |
-
causal=False,
|
148 |
-
softmax_scale=None,
|
149 |
-
attention_dropout=0.0,
|
150 |
-
):
|
151 |
-
super().__init__()
|
152 |
-
assert flash_attn_unpadded_func is not None, (
|
153 |
-
"Please install FlashAttention first, " "e.g., with pip install flash-attn"
|
154 |
-
)
|
155 |
-
assert (
|
156 |
-
rearrange is not None
|
157 |
-
), "Please install einops first, e.g., with pip install einops"
|
158 |
-
self.causal = causal
|
159 |
-
self.softmax_scale = softmax_scale
|
160 |
-
self.dropout_p = attention_dropout
|
161 |
-
|
162 |
-
def forward(self, q, k, v):
|
163 |
-
assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
|
164 |
-
assert all((i.is_cuda for i in (q, k, v)))
|
165 |
-
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
166 |
-
seqlen_k = k.shape[1]
|
167 |
-
q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
|
168 |
-
cu_seqlens_q = torch.arange(
|
169 |
-
0,
|
170 |
-
(batch_size + 1) * seqlen_q,
|
171 |
-
step=seqlen_q,
|
172 |
-
dtype=torch.int32,
|
173 |
-
device=q.device,
|
174 |
-
)
|
175 |
-
|
176 |
-
if self.training:
|
177 |
-
assert seqlen_k == seqlen_q
|
178 |
-
|
179 |
-
is_causal = self.causal
|
180 |
-
cu_seqlens_k = cu_seqlens_q
|
181 |
-
else:
|
182 |
-
is_causal = seqlen_q == seqlen_k
|
183 |
-
cu_seqlens_k = torch.arange(
|
184 |
-
0,
|
185 |
-
(batch_size + 1) * seqlen_k,
|
186 |
-
step=seqlen_k,
|
187 |
-
dtype=torch.int32,
|
188 |
-
device=q.device,
|
189 |
-
)
|
190 |
-
self.dropout_p = 0
|
191 |
-
output = flash_attn_unpadded_func(
|
192 |
-
q,
|
193 |
-
k,
|
194 |
-
v,
|
195 |
-
cu_seqlens_q,
|
196 |
-
cu_seqlens_k,
|
197 |
-
seqlen_q,
|
198 |
-
seqlen_k,
|
199 |
-
self.dropout_p,
|
200 |
-
softmax_scale=self.softmax_scale,
|
201 |
-
causal=is_causal,
|
202 |
-
)
|
203 |
-
|
204 |
-
output = rearrange(output, "(b s) ... -> b s ...", b=batch_size)
|
205 |
-
return output
|
206 |
-
|
207 |
-
|
208 |
class QWenAttention(nn.Module):
|
209 |
def __init__(self, config):
|
210 |
super().__init__()
|
@@ -225,7 +124,6 @@ class QWenAttention(nn.Module):
|
|
225 |
self.num_heads = config.num_attention_heads
|
226 |
self.head_dim = self.hidden_size // self.num_heads
|
227 |
|
228 |
-
self.use_flash_attn = config.use_flash_attn
|
229 |
self.scale_attn_weights = True
|
230 |
|
231 |
self.projection_size = config.kv_channels * config.num_attention_heads
|
@@ -242,15 +140,6 @@ class QWenAttention(nn.Module):
|
|
242 |
)
|
243 |
|
244 |
self.is_fp32 = not (config.bf16 or config.fp16)
|
245 |
-
if (
|
246 |
-
self.use_flash_attn
|
247 |
-
and flash_attn_unpadded_func is not None
|
248 |
-
and not self.is_fp32
|
249 |
-
):
|
250 |
-
self.core_attention_flash = FlashSelfAttention(
|
251 |
-
causal=True, attention_dropout=config.attn_dropout_prob
|
252 |
-
)
|
253 |
-
|
254 |
self.bf16 = config.bf16
|
255 |
|
256 |
if config.rotary_pct == 1.0:
|
@@ -453,40 +342,20 @@ class QWenAttention(nn.Module):
|
|
453 |
logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
|
454 |
query = query * logn_tensor.expand_as(query)
|
455 |
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
)
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
context_layer = rearrange(
|
466 |
-
context_layer, "b s h d -> b s (h d)"
|
467 |
-
).contiguous()
|
468 |
-
else:
|
469 |
-
query = query.permute(0, 2, 1, 3)
|
470 |
-
key = key.permute(0, 2, 1, 3)
|
471 |
-
value = value.permute(0, 2, 1, 3)
|
472 |
-
attn_output, attn_weight = self._attn(
|
473 |
-
query, key, value, attention_mask, head_mask
|
474 |
-
)
|
475 |
-
context_layer = self._merge_heads(
|
476 |
-
attn_output, self.num_heads, self.head_dim
|
477 |
-
)
|
478 |
|
479 |
attn_output = self.c_proj(context_layer)
|
480 |
outputs = (attn_output, present)
|
481 |
if output_attentions:
|
482 |
-
|
483 |
-
self.use_flash_attn
|
484 |
-
and flash_attn_unpadded_func is not None
|
485 |
-
and not self.is_fp32
|
486 |
-
):
|
487 |
-
raise ValueError("Cannot output attentions while using flash-attn")
|
488 |
-
else:
|
489 |
-
outputs += (attn_weight,)
|
490 |
|
491 |
return outputs
|
492 |
|
@@ -882,18 +751,6 @@ class QWenLMHeadModel(QWenPreTrainedModel):
|
|
882 |
logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
|
883 |
elif SUPPORT_FP16:
|
884 |
logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
|
885 |
-
|
886 |
-
if config.use_flash_attn == "auto":
|
887 |
-
if config.bf16 or config.fp16:
|
888 |
-
logger.warn("Try importing flash-attention for faster inference...")
|
889 |
-
config.use_flash_attn = True
|
890 |
-
else:
|
891 |
-
config.use_flash_attn = False
|
892 |
-
if config.use_flash_attn and config.fp32:
|
893 |
-
logger.warn("Flash attention will be disabled because it does NOT support fp32.")
|
894 |
-
|
895 |
-
if config.use_flash_attn:
|
896 |
-
_import_flash_attn()
|
897 |
|
898 |
self.transformer = QWenModel(config)
|
899 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
|
|
69 |
|
70 |
apply_rotary_emb_func = None
|
71 |
rms_norm = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
75 |
def _make_causal_mask(
|
|
|
104 |
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
105 |
|
106 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
class QWenAttention(nn.Module):
|
108 |
def __init__(self, config):
|
109 |
super().__init__()
|
|
|
124 |
self.num_heads = config.num_attention_heads
|
125 |
self.head_dim = self.hidden_size // self.num_heads
|
126 |
|
|
|
127 |
self.scale_attn_weights = True
|
128 |
|
129 |
self.projection_size = config.kv_channels * config.num_attention_heads
|
|
|
140 |
)
|
141 |
|
142 |
self.is_fp32 = not (config.bf16 or config.fp16)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
143 |
self.bf16 = config.bf16
|
144 |
|
145 |
if config.rotary_pct == 1.0:
|
|
|
342 |
logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
|
343 |
query = query * logn_tensor.expand_as(query)
|
344 |
|
345 |
+
query = query.permute(0, 2, 1, 3)
|
346 |
+
key = key.permute(0, 2, 1, 3)
|
347 |
+
value = value.permute(0, 2, 1, 3)
|
348 |
+
attn_output, attn_weight = self._attn(
|
349 |
+
query, key, value, attention_mask, head_mask
|
350 |
+
)
|
351 |
+
context_layer = self._merge_heads(
|
352 |
+
attn_output, self.num_heads, self.head_dim
|
353 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
354 |
|
355 |
attn_output = self.c_proj(context_layer)
|
356 |
outputs = (attn_output, present)
|
357 |
if output_attentions:
|
358 |
+
outputs += (attn_weight,)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
359 |
|
360 |
return outputs
|
361 |
|
|
|
751 |
logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
|
752 |
elif SUPPORT_FP16:
|
753 |
logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
754 |
|
755 |
self.transformer = QWenModel(config)
|
756 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
tokenization_qwen.py
CHANGED
@@ -10,7 +10,7 @@ import logging
|
|
10 |
import os
|
11 |
import requests
|
12 |
import unicodedata
|
13 |
-
from typing import Collection, Dict, List, Set, Tuple, Union, Any, Callable
|
14 |
|
15 |
import tiktoken
|
16 |
import numpy as np
|
@@ -359,6 +359,22 @@ class QWenTokenizer(PreTrainedTokenizer):
|
|
359 |
_encode_vl_info,
|
360 |
)
|
361 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
362 |
def _fetch_latest_picture(self, response, history):
|
363 |
if history is None:
|
364 |
history = []
|
@@ -377,15 +393,19 @@ class QWenTokenizer(PreTrainedTokenizer):
|
|
377 |
bbox = tuple(map(int, ele['box'].replace('(', '').replace(')', '').split(',')))
|
378 |
assert len(bbox) == 4
|
379 |
output.append({'box': bbox})
|
380 |
-
|
381 |
-
|
|
|
|
|
|
|
|
|
382 |
return output
|
383 |
|
384 |
def draw_bbox_on_latest_picture(
|
385 |
self,
|
386 |
response,
|
387 |
history=None,
|
388 |
-
):
|
389 |
image = self._fetch_latest_picture(response, history)
|
390 |
if image is None:
|
391 |
return None
|
@@ -399,14 +419,14 @@ class QWenTokenizer(PreTrainedTokenizer):
|
|
399 |
boxes = self._fetch_all_box_with_ref(response)
|
400 |
if not boxes:
|
401 |
return None
|
402 |
-
fnt = ImageFont.truetype("SimSun.ttf",
|
403 |
draw = ImageDraw.Draw(image)
|
404 |
for box in boxes:
|
405 |
x1, y1, x2, y2 = box['box']
|
406 |
x1, y1, x2, y2 = (int(x1 / 1000 * w), int(y1 / 1000 * h), int(x2 / 1000 * w), int(y2 / 1000 * h))
|
407 |
-
draw.rectangle((x1, y1, x2, y2), outline='red', width=
|
408 |
if 'ref' in box:
|
409 |
-
draw.text((x1, y1), box['ref'], fill='
|
410 |
return image
|
411 |
|
412 |
|
|
|
10 |
import os
|
11 |
import requests
|
12 |
import unicodedata
|
13 |
+
from typing import Collection, Dict, List, Set, Tuple, Union, Any, Callable, Optional
|
14 |
|
15 |
import tiktoken
|
16 |
import numpy as np
|
|
|
359 |
_encode_vl_info,
|
360 |
)
|
361 |
|
362 |
+
def from_list_format(self, list_format: List[Dict]):
|
363 |
+
text = ''
|
364 |
+
for ele in list_format:
|
365 |
+
if 'image' in ele:
|
366 |
+
text += self.image_start_tag + ele['image'] + self.image_end_tag
|
367 |
+
elif 'text' in ele:
|
368 |
+
text += ele['text']
|
369 |
+
elif 'box' in ele:
|
370 |
+
if 'ref' in ele:
|
371 |
+
text += self.ref_start_tag + ele['ref'] + self.ref_end_tag
|
372 |
+
for box in ele['box']:
|
373 |
+
text += self.box_start_tag + '(%d,%d),(%d,%d)' % (box[0], box[1], box[2], box[3]) + self.box_end_tag
|
374 |
+
else:
|
375 |
+
raise ValueError("Unsupport element: " + str(ele))
|
376 |
+
return text
|
377 |
+
|
378 |
def _fetch_latest_picture(self, response, history):
|
379 |
if history is None:
|
380 |
history = []
|
|
|
393 |
bbox = tuple(map(int, ele['box'].replace('(', '').replace(')', '').split(',')))
|
394 |
assert len(bbox) == 4
|
395 |
output.append({'box': bbox})
|
396 |
+
|
397 |
+
ref_idx = i - 1
|
398 |
+
while ref_idx >= 0 and 'box' in list_format[ref_idx]:
|
399 |
+
ref_idx -= 1
|
400 |
+
if ref_idx >= 0 and 'ref' in list_format[ref_idx]:
|
401 |
+
output[-1]['ref'] = list_format[ref_idx]['ref'].strip()
|
402 |
return output
|
403 |
|
404 |
def draw_bbox_on_latest_picture(
|
405 |
self,
|
406 |
response,
|
407 |
history=None,
|
408 |
+
) -> Optional[Image.Image]:
|
409 |
image = self._fetch_latest_picture(response, history)
|
410 |
if image is None:
|
411 |
return None
|
|
|
419 |
boxes = self._fetch_all_box_with_ref(response)
|
420 |
if not boxes:
|
421 |
return None
|
422 |
+
fnt = ImageFont.truetype("SimSun.ttf", 50)
|
423 |
draw = ImageDraw.Draw(image)
|
424 |
for box in boxes:
|
425 |
x1, y1, x2, y2 = box['box']
|
426 |
x1, y1, x2, y2 = (int(x1 / 1000 * w), int(y1 / 1000 * h), int(x2 / 1000 * w), int(y2 / 1000 * h))
|
427 |
+
draw.rectangle((x1, y1, x2, y2), outline='red', width=4)
|
428 |
if 'ref' in box:
|
429 |
+
draw.text((x1, y1), box['ref'], fill='yellow', font=fnt)
|
430 |
return image
|
431 |
|
432 |
|
visual.py
CHANGED
@@ -1,3 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from collections import OrderedDict
|
2 |
import math
|
3 |
import requests
|
@@ -5,11 +10,11 @@ from io import BytesIO
|
|
5 |
from functools import partial
|
6 |
from PIL import Image
|
7 |
from typing import Callable, Optional, Sequence, Tuple, List
|
|
|
8 |
|
9 |
import torch
|
10 |
from torch import nn
|
11 |
from torch.nn import functional as F
|
12 |
-
from torch.utils.checkpoint import checkpoint
|
13 |
from torch.nn.init import trunc_normal_
|
14 |
from torchvision import transforms
|
15 |
from torchvision.transforms import InterpolationMode
|
@@ -33,8 +38,64 @@ def get_abs_pos(abs_pos, tgt_size):
|
|
33 |
else:
|
34 |
return abs_pos
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
class Resampler(nn.Module):
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
def __init__(
|
39 |
self,
|
40 |
grid_size,
|
@@ -48,7 +109,9 @@ class Resampler(nn.Module):
|
|
48 |
self.embed_dim = embed_dim
|
49 |
self.num_heads = num_heads
|
50 |
|
51 |
-
self.pos_embed = nn.Parameter(
|
|
|
|
|
52 |
|
53 |
self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
|
54 |
trunc_normal_(self.query, std=.02)
|
@@ -234,7 +297,7 @@ class VisualAttentionBlock(nn.Module):
|
|
234 |
return x
|
235 |
|
236 |
|
237 |
-
class
|
238 |
def __init__(
|
239 |
self,
|
240 |
width: int,
|
@@ -247,7 +310,6 @@ class Transformer(nn.Module):
|
|
247 |
super().__init__()
|
248 |
self.width = width
|
249 |
self.layers = layers
|
250 |
-
self.grad_checkpointing = False
|
251 |
|
252 |
self.resblocks = nn.ModuleList([
|
253 |
VisualAttentionBlock(
|
@@ -263,11 +325,7 @@ class Transformer(nn.Module):
|
|
263 |
|
264 |
def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
|
265 |
for r in self.resblocks:
|
266 |
-
|
267 |
-
# TODO: handle kwargs https://github.com/pytorch/pytorch/issues/79887#issuecomment-1161758372
|
268 |
-
x = checkpoint(r, x, None, None, attn_mask)
|
269 |
-
else:
|
270 |
-
x = r(x, attn_mask=attn_mask)
|
271 |
return x
|
272 |
|
273 |
|
@@ -306,13 +364,13 @@ class VisionTransformer(nn.Module):
|
|
306 |
|
307 |
# class embeddings and positional embeddings
|
308 |
scale = width ** -0.5
|
309 |
-
self.positional_embedding = nn.Parameter(scale * torch.randn(
|
310 |
|
311 |
norm_layer = partial(nn.LayerNorm, eps=1e-6)
|
312 |
act_layer = nn.GELU
|
313 |
|
314 |
self.ln_pre = norm_layer(width)
|
315 |
-
self.transformer =
|
316 |
width,
|
317 |
layers,
|
318 |
heads,
|
@@ -331,10 +389,6 @@ class VisionTransformer(nn.Module):
|
|
331 |
self.ln_post = norm_layer(output_dim)
|
332 |
self.proj = nn.Parameter((output_dim** -0.5) * torch.randn(output_dim, output_dim))
|
333 |
|
334 |
-
@torch.jit.ignore
|
335 |
-
def set_grad_checkpointing(self, enable=True):
|
336 |
-
self.transformer.grad_checkpointing = enable
|
337 |
-
|
338 |
def forward(self, x: torch.Tensor):
|
339 |
x = x.to(
|
340 |
dtype=self.transformer.get_cast_dtype(),
|
@@ -353,8 +407,7 @@ class VisionTransformer(nn.Module):
|
|
353 |
x = self.transformer(x)
|
354 |
x = x.permute(1, 0, 2) # LND -> NLD
|
355 |
|
356 |
-
|
357 |
-
x = self.attn_pool(x)
|
358 |
x = self.ln_post(x)
|
359 |
x = x @ self.proj
|
360 |
|
@@ -365,8 +418,6 @@ class VisionTransformer(nn.Module):
|
|
365 |
for image_path in image_paths:
|
366 |
if image_path.startswith("http://") or image_path.startswith("https://"):
|
367 |
image = Image.open(requests.get(image_path, stream=True).raw)
|
368 |
-
elif image_path.startswith("oss://"):
|
369 |
-
raise NotImplementedError
|
370 |
else:
|
371 |
image = Image.open(image_path)
|
372 |
image = image.convert("RGB")
|
|
|
1 |
+
# Copyright (c) Alibaba Cloud.
|
2 |
+
#
|
3 |
+
# This source code is licensed under the license found in the
|
4 |
+
# LICENSE file in the root directory of this source tree.
|
5 |
+
|
6 |
from collections import OrderedDict
|
7 |
import math
|
8 |
import requests
|
|
|
10 |
from functools import partial
|
11 |
from PIL import Image
|
12 |
from typing import Callable, Optional, Sequence, Tuple, List
|
13 |
+
import numpy as np
|
14 |
|
15 |
import torch
|
16 |
from torch import nn
|
17 |
from torch.nn import functional as F
|
|
|
18 |
from torch.nn.init import trunc_normal_
|
19 |
from torchvision import transforms
|
20 |
from torchvision.transforms import InterpolationMode
|
|
|
38 |
else:
|
39 |
return abs_pos
|
40 |
|
41 |
+
# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
|
42 |
+
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
|
43 |
+
"""
|
44 |
+
grid_size: int of the grid height and width
|
45 |
+
return:
|
46 |
+
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
|
47 |
+
"""
|
48 |
+
grid_h = np.arange(grid_size, dtype=np.float32)
|
49 |
+
grid_w = np.arange(grid_size, dtype=np.float32)
|
50 |
+
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
51 |
+
grid = np.stack(grid, axis=0)
|
52 |
+
|
53 |
+
grid = grid.reshape([2, 1, grid_size, grid_size])
|
54 |
+
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
55 |
+
if cls_token:
|
56 |
+
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
|
57 |
+
return pos_embed
|
58 |
+
|
59 |
+
|
60 |
+
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
|
61 |
+
assert embed_dim % 2 == 0
|
62 |
+
|
63 |
+
# use half of dimensions to encode grid_h
|
64 |
+
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
|
65 |
+
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
|
66 |
+
|
67 |
+
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
|
68 |
+
return emb
|
69 |
+
|
70 |
+
|
71 |
+
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
|
72 |
+
"""
|
73 |
+
embed_dim: output dimension for each position
|
74 |
+
pos: a list of positions to be encoded: size (M,)
|
75 |
+
out: (M, D)
|
76 |
+
"""
|
77 |
+
assert embed_dim % 2 == 0
|
78 |
+
omega = np.arange(embed_dim // 2, dtype=np.float32)
|
79 |
+
omega /= embed_dim / 2.
|
80 |
+
omega = 1. / 10000**omega # (D/2,)
|
81 |
+
|
82 |
+
pos = pos.reshape(-1) # (M,)
|
83 |
+
out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
|
84 |
+
|
85 |
+
emb_sin = np.sin(out) # (M, D/2)
|
86 |
+
emb_cos = np.cos(out) # (M, D/2)
|
87 |
+
|
88 |
+
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
|
89 |
+
return emb
|
90 |
+
|
91 |
|
92 |
class Resampler(nn.Module):
|
93 |
+
"""
|
94 |
+
A 2D perceiver-resampler network with one cross attention layers by
|
95 |
+
(grid_size**2) learnable queries and 2d sincos pos_emb
|
96 |
+
Outputs:
|
97 |
+
A tensor with the shape of (grid_size**2, embed_dim)
|
98 |
+
"""
|
99 |
def __init__(
|
100 |
self,
|
101 |
grid_size,
|
|
|
109 |
self.embed_dim = embed_dim
|
110 |
self.num_heads = num_heads
|
111 |
|
112 |
+
self.pos_embed = nn.Parameter(
|
113 |
+
torch.from_numpy(get_2d_sincos_pos_embed(embed_dim, grid_size)).float()
|
114 |
+
).requires_grad_(False)
|
115 |
|
116 |
self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
|
117 |
trunc_normal_(self.query, std=.02)
|
|
|
297 |
return x
|
298 |
|
299 |
|
300 |
+
class TransformerBlock(nn.Module):
|
301 |
def __init__(
|
302 |
self,
|
303 |
width: int,
|
|
|
310 |
super().__init__()
|
311 |
self.width = width
|
312 |
self.layers = layers
|
|
|
313 |
|
314 |
self.resblocks = nn.ModuleList([
|
315 |
VisualAttentionBlock(
|
|
|
325 |
|
326 |
def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
|
327 |
for r in self.resblocks:
|
328 |
+
x = r(x, attn_mask=attn_mask)
|
|
|
|
|
|
|
|
|
329 |
return x
|
330 |
|
331 |
|
|
|
364 |
|
365 |
# class embeddings and positional embeddings
|
366 |
scale = width ** -0.5
|
367 |
+
self.positional_embedding = nn.Parameter(scale * torch.randn(256, width))
|
368 |
|
369 |
norm_layer = partial(nn.LayerNorm, eps=1e-6)
|
370 |
act_layer = nn.GELU
|
371 |
|
372 |
self.ln_pre = norm_layer(width)
|
373 |
+
self.transformer = TransformerBlock(
|
374 |
width,
|
375 |
layers,
|
376 |
heads,
|
|
|
389 |
self.ln_post = norm_layer(output_dim)
|
390 |
self.proj = nn.Parameter((output_dim** -0.5) * torch.randn(output_dim, output_dim))
|
391 |
|
|
|
|
|
|
|
|
|
392 |
def forward(self, x: torch.Tensor):
|
393 |
x = x.to(
|
394 |
dtype=self.transformer.get_cast_dtype(),
|
|
|
407 |
x = self.transformer(x)
|
408 |
x = x.permute(1, 0, 2) # LND -> NLD
|
409 |
|
410 |
+
x = self.attn_pool(x)
|
|
|
411 |
x = self.ln_post(x)
|
412 |
x = x @ self.proj
|
413 |
|
|
|
418 |
for image_path in image_paths:
|
419 |
if image_path.startswith("http://") or image_path.startswith("https://"):
|
420 |
image = Image.open(requests.get(image_path, stream=True).raw)
|
|
|
|
|
421 |
else:
|
422 |
image = Image.open(image_path)
|
423 |
image = image.convert("RGB")
|