Create app-multi-alpha.py
Browse files- app-multi-alpha.py +210 -0
app-multi-alpha.py
ADDED
@@ -0,0 +1,210 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.amp.autocast_mode
|
3 |
+
import torch.distributed as dist
|
4 |
+
import torch.multiprocessing as mp
|
5 |
+
from torch.nn.parallel import DistributedDataParallel as DDP
|
6 |
+
import os
|
7 |
+
import sys
|
8 |
+
import logging
|
9 |
+
import warnings
|
10 |
+
import argparse
|
11 |
+
from PIL import Image
|
12 |
+
from pathlib import Path
|
13 |
+
from tqdm import tqdm
|
14 |
+
from torch import nn
|
15 |
+
from transformers import AutoModel, AutoProcessor, AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast, AutoModelForCausalLM
|
16 |
+
from typing import List, Union
|
17 |
+
|
18 |
+
# Constants
|
19 |
+
CLIP_PATH = "google/siglip-so400m-patch14-384"
|
20 |
+
VLM_PROMPT = "A descriptive caption for this image:\n"
|
21 |
+
MODEL_PATH = "unsloth/Meta-Llama-3.1-8B-bnb-4bit"
|
22 |
+
CHECKPOINT_PATH = Path("wpkklhc6")
|
23 |
+
IMAGE_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.bmp', '.webp')
|
24 |
+
|
25 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
26 |
+
logging.getLogger("transformers").setLevel(logging.ERROR)
|
27 |
+
|
28 |
+
def setup(rank, world_size):
|
29 |
+
os.environ['MASTER_ADDR'] = 'localhost'
|
30 |
+
os.environ['MASTER_PORT'] = '12355'
|
31 |
+
dist.init_process_group("nccl", rank=rank, world_size=world_size)
|
32 |
+
|
33 |
+
def cleanup():
|
34 |
+
dist.destroy_process_group()
|
35 |
+
|
36 |
+
class ImageAdapter(nn.Module):
|
37 |
+
def __init__(self, input_features: int, output_features: int):
|
38 |
+
super().__init__()
|
39 |
+
self.linear1 = nn.Linear(input_features, output_features)
|
40 |
+
self.activation = nn.GELU()
|
41 |
+
self.linear2 = nn.Linear(output_features, output_features)
|
42 |
+
|
43 |
+
def forward(self, vision_outputs: torch.Tensor):
|
44 |
+
return self.linear2(self.activation(self.linear1(vision_outputs)))
|
45 |
+
|
46 |
+
def load_models(rank):
|
47 |
+
print(f"Loading CLIP π on GPU {rank}")
|
48 |
+
clip_processor = AutoProcessor.from_pretrained(CLIP_PATH)
|
49 |
+
clip_model = AutoModel.from_pretrained(CLIP_PATH).vision_model.eval().requires_grad_(False).to(rank)
|
50 |
+
|
51 |
+
print(f"Loading tokenizer πͺ on GPU {rank}")
|
52 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, use_fast=False)
|
53 |
+
assert isinstance(tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast)), f"Tokenizer is of type {type(tokenizer)}"
|
54 |
+
|
55 |
+
print(f"Loading LLM π€ on GPU {rank}")
|
56 |
+
text_model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, device_map={"": rank}, torch_dtype=torch.bfloat16).eval()
|
57 |
+
|
58 |
+
print(f"Loading image adapter πΌοΈ on GPU {rank}")
|
59 |
+
image_adapter = ImageAdapter(clip_model.config.hidden_size, text_model.config.hidden_size)
|
60 |
+
image_adapter.load_state_dict(torch.load(CHECKPOINT_PATH / "image_adapter.pt", map_location=f"cuda:{rank}", weights_only=True))
|
61 |
+
image_adapter.eval().to(rank)
|
62 |
+
|
63 |
+
return clip_processor, clip_model, tokenizer, text_model, image_adapter
|
64 |
+
|
65 |
+
@torch.no_grad()
|
66 |
+
def stream_chat(input_images: List[Image.Image], batch_size: int, pbar: tqdm, models: tuple, rank: int) -> List[str]:
|
67 |
+
clip_processor, clip_model, tokenizer, text_model, image_adapter = models
|
68 |
+
torch.cuda.empty_cache()
|
69 |
+
all_captions = []
|
70 |
+
|
71 |
+
for i in range(0, len(input_images), batch_size):
|
72 |
+
batch = input_images[i:i+batch_size]
|
73 |
+
|
74 |
+
try:
|
75 |
+
images = clip_processor(images=batch, return_tensors='pt', padding=True).pixel_values.to(rank)
|
76 |
+
except ValueError as e:
|
77 |
+
print(f"Error processing image batch: {e}")
|
78 |
+
print("Skipping this batch and continuing...")
|
79 |
+
continue
|
80 |
+
|
81 |
+
with torch.amp.autocast_mode.autocast(rank, enabled=True):
|
82 |
+
vision_outputs = clip_model(pixel_values=images, output_hidden_states=True)
|
83 |
+
image_features = vision_outputs.hidden_states[-2]
|
84 |
+
embedded_images = image_adapter(image_features).to(dtype=torch.bfloat16)
|
85 |
+
|
86 |
+
prompt = tokenizer.encode(VLM_PROMPT, return_tensors='pt')
|
87 |
+
prompt_embeds = text_model.model.embed_tokens(prompt.to(rank)).to(dtype=torch.bfloat16)
|
88 |
+
embedded_bos = text_model.model.embed_tokens(torch.tensor([[tokenizer.bos_token_id]], device=rank, dtype=torch.int64)).to(dtype=torch.bfloat16)
|
89 |
+
|
90 |
+
inputs_embeds = torch.cat([
|
91 |
+
embedded_bos.expand(embedded_images.shape[0], -1, -1),
|
92 |
+
embedded_images,
|
93 |
+
prompt_embeds.expand(embedded_images.shape[0], -1, -1),
|
94 |
+
], dim=1).to(dtype=torch.bfloat16)
|
95 |
+
|
96 |
+
input_ids = torch.cat([
|
97 |
+
torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long).expand(embedded_images.shape[0], -1),
|
98 |
+
torch.zeros((embedded_images.shape[0], embedded_images.shape[1]), dtype=torch.long),
|
99 |
+
prompt.expand(embedded_images.shape[0], -1),
|
100 |
+
], dim=1).to(rank)
|
101 |
+
|
102 |
+
attention_mask = torch.ones_like(input_ids)
|
103 |
+
|
104 |
+
generate_ids = text_model.generate(
|
105 |
+
input_ids=input_ids,
|
106 |
+
inputs_embeds=inputs_embeds,
|
107 |
+
attention_mask=attention_mask,
|
108 |
+
max_new_tokens=300,
|
109 |
+
do_sample=True,
|
110 |
+
top_k=10,
|
111 |
+
temperature=0.5,
|
112 |
+
)
|
113 |
+
|
114 |
+
generate_ids = generate_ids[:, input_ids.shape[1]:]
|
115 |
+
|
116 |
+
for ids in generate_ids:
|
117 |
+
caption = tokenizer.decode(ids[:-1] if ids[-1] == tokenizer.eos_token_id else ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
|
118 |
+
caption = caption.replace('<|end_of_text|>', '').replace('<|finetune_right_pad_id|>', '').strip()
|
119 |
+
all_captions.append(caption)
|
120 |
+
|
121 |
+
if pbar and rank == 0:
|
122 |
+
pbar.update(len(batch))
|
123 |
+
|
124 |
+
return all_captions
|
125 |
+
|
126 |
+
def process_directory(rank, world_size, input_dir: Path, output_dir: Path, batch_size: int, models: tuple):
|
127 |
+
output_dir.mkdir(parents=True, exist_ok=True)
|
128 |
+
image_files = [f for f in input_dir.iterdir() if f.suffix.lower() in IMAGE_EXTENSIONS]
|
129 |
+
images_to_process = [f for f in image_files if not (output_dir / f"{f.stem}.txt").exists()]
|
130 |
+
|
131 |
+
if not images_to_process:
|
132 |
+
if rank == 0:
|
133 |
+
print("No new images to process.")
|
134 |
+
return
|
135 |
+
|
136 |
+
# Distribute images across GPUs
|
137 |
+
images_per_gpu = len(images_to_process) // world_size
|
138 |
+
start_idx = rank * images_per_gpu
|
139 |
+
end_idx = start_idx + images_per_gpu if rank < world_size - 1 else len(images_to_process)
|
140 |
+
gpu_images = images_to_process[start_idx:end_idx]
|
141 |
+
|
142 |
+
if rank == 0:
|
143 |
+
pbar = tqdm(total=len(images_to_process), desc="Processing images", unit="image")
|
144 |
+
else:
|
145 |
+
pbar = None
|
146 |
+
|
147 |
+
for i in range(0, len(gpu_images), batch_size):
|
148 |
+
batch_files = gpu_images[i:i+batch_size]
|
149 |
+
batch_images = [Image.open(f).convert('RGB') for f in batch_files]
|
150 |
+
|
151 |
+
captions = stream_chat(batch_images, batch_size, pbar, models, rank)
|
152 |
+
|
153 |
+
for file, caption in zip(batch_files, captions):
|
154 |
+
with open(output_dir / f"{file.stem}.txt", 'w', encoding='utf-8') as f:
|
155 |
+
f.write(caption)
|
156 |
+
|
157 |
+
for img in batch_images:
|
158 |
+
img.close()
|
159 |
+
|
160 |
+
if rank == 0:
|
161 |
+
pbar.close()
|
162 |
+
|
163 |
+
def parse_arguments():
|
164 |
+
parser = argparse.ArgumentParser(description="Process images and generate captions.")
|
165 |
+
parser.add_argument("input", nargs='+', help="Input image file or directory (or multiple directories)")
|
166 |
+
parser.add_argument("--output", help="Output directory (optional)")
|
167 |
+
parser.add_argument("--bs", type=int, default=4, help="Batch size (default: 4)")
|
168 |
+
return parser.parse_args()
|
169 |
+
|
170 |
+
def run(rank, world_size, args):
|
171 |
+
setup(rank, world_size)
|
172 |
+
|
173 |
+
input_paths = [Path(input_path) for input_path in args.input]
|
174 |
+
batch_size = args.bs
|
175 |
+
models = load_models(rank)
|
176 |
+
|
177 |
+
for input_path in input_paths:
|
178 |
+
if input_path.is_file() and input_path.suffix.lower() in IMAGE_EXTENSIONS:
|
179 |
+
if rank == 0:
|
180 |
+
output_path = input_path.with_suffix('.txt')
|
181 |
+
print(f"Processing single image ποΈ: {input_path.name}")
|
182 |
+
with tqdm(total=1, desc="Processing image", unit="image") as pbar:
|
183 |
+
captions = stream_chat([Image.open(input_path).convert('RGB')], 1, pbar, models, rank)
|
184 |
+
with open(output_path, 'w', encoding='utf-8') as f:
|
185 |
+
f.write(captions[0])
|
186 |
+
print(f"Output saved to {output_path}")
|
187 |
+
elif input_path.is_dir():
|
188 |
+
output_path = Path(args.output) if args.output else input_path
|
189 |
+
if rank == 0:
|
190 |
+
print(f"Processing directory π: {input_path}")
|
191 |
+
print(f"Output directory π¦: {output_path}")
|
192 |
+
print(f"Batch size ποΈ: {batch_size}")
|
193 |
+
process_directory(rank, world_size, input_path, output_path, batch_size, models)
|
194 |
+
else:
|
195 |
+
if rank == 0:
|
196 |
+
print(f"Invalid input: {input_path}")
|
197 |
+
print("Skipping...")
|
198 |
+
|
199 |
+
cleanup()
|
200 |
+
|
201 |
+
def main():
|
202 |
+
args = parse_arguments()
|
203 |
+
world_size = torch.cuda.device_count()
|
204 |
+
if world_size > 1:
|
205 |
+
mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
|
206 |
+
else:
|
207 |
+
run(0, 1, args)
|
208 |
+
|
209 |
+
if __name__ == "__main__":
|
210 |
+
main()
|