File size: 8,541 Bytes
95f4d99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc844d3
 
 
 
 
 
 
95f4d99
 
 
 
d6dda21
95f4d99
 
 
 
cc844d3
 
 
 
 
 
 
95f4d99
cc844d3
 
 
 
 
95f4d99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a6c7d5
 
 
95f4d99
cc844d3
 
95f4d99
 
 
 
cc844d3
95f4d99
 
 
 
 
 
cc844d3
 
95f4d99
 
 
 
 
 
 
 
cc844d3
 
 
bc0c9d8
95f4d99
 
 
 
 
 
 
 
 
 
 
 
 
 
cc844d3
 
0c504eb
95f4d99
 
 
 
 
0c504eb
95f4d99
b94a4c6
 
95f4d99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d6dda21
95f4d99
 
 
d6dda21
 
 
 
 
 
 
95f4d99
 
 
 
cc844d3
95f4d99
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
# # main.py
# from fastapi import FastAPI, File, UploadFile
# from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration
# from transformers.image_utils import load_image
# import torch
# from io import BytesIO
# import os
# from dotenv import load_dotenv
# from PIL import Image

# from huggingface_hub import login

# # Load environment variables
# load_dotenv()

# # Set the cache directory to a writable path
# os.environ["TORCHINDUCTOR_CACHE_DIR"] = "/tmp/torch_inductor_cache"

# token = os.getenv("huggingface_ankit")
# # Login to the Hugging Face Hub
# login(token)

# app = FastAPI()

# model_id = "google/paligemma2-3b-mix-448"
# model = PaliGemmaForConditionalGeneration.from_pretrained(model_id).to('cuda')
# processor = PaliGemmaProcessor.from_pretrained(model_id)

# def predict(image):
#     prompt = "<image> ocr"
#     model_inputs = processor(text=prompt, images=image, return_tensors="pt").to('cuda')
#     input_len = model_inputs["input_ids"].shape[-1]
#     with torch.inference_mode():
#         generation = model.generate(**model_inputs, max_new_tokens=200)
#     torch.cuda.empty_cache()
#     decoded = processor.decode(generation[0], skip_special_tokens=True) #[len(prompt):].lstrip("\n")
#     return decoded

# @app.post("/extract_text")
# async def extract_text(file: UploadFile = File(...)):
#     image = Image.open(BytesIO(await file.read())).convert("RGB")  # Ensure it's a valid PIL image
#     text = predict(image)
#     return {"extracted_text": text}

# @app.post("/batch_extract_text")
# async def batch_extract_text(files: list[UploadFile] = File(...)):
#     # if len(files) > 20:
#     #     return {"error": "A maximum of 20 images can be processed at a time."}
    
#     images = [Image.open(BytesIO(await file.read())).convert("RGB") for file in files]
#     prompts = ["OCR"] * len(images)
    
#     model_inputs = processor(text=prompts, images=images, return_tensors="pt").to(torch.bfloat16).to(model.device)
#     input_len = model_inputs["input_ids"].shape[-1]
    
#     with torch.inference_mode():
#         generations = model.generate(**model_inputs, max_new_tokens=200, do_sample=False)
#     torch.cuda.empty_cache()
#     extracted_texts = [processor.decode(generations[i], skip_special_tokens=True) for i in range(len(images))]
    
#     return {"extracted_texts": extracted_texts}
    
# if __name__ == "__main__":
#     import uvicorn
#     uvicorn.run(app, host="0.0.0.0", port=7860)

from fastapi import FastAPI, File, UploadFile, BackgroundTasks
from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration
import torch
from io import BytesIO
import os
from dotenv import load_dotenv
from PIL import Image
from huggingface_hub import login
import gc
import logging
from typing import List
import time
import numpy as np

# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# Load environment variables
load_dotenv()

# Set the cache directory to a writable path
os.environ["TORCHINDUCTOR_CACHE_DIR"] = "/tmp/torch_inductor_cache"
token = os.getenv("huggingface_ankit")

# Login to the Hugging Face Hub
login(token)

app = FastAPI()

# Global variables for model and processor
model = None
processor = None

def load_model():
    """Load model and processor when needed"""
    global model, processor
    if model is None:
        model_id = "google/paligemma2-3b-mix-448"
        logger.info(f"Loading model {model_id}")
        
        # Load model with memory-efficient settings
        model = PaliGemmaForConditionalGeneration.from_pretrained(
            model_id,
            device_map="auto",
            torch_dtype=torch.bfloat16  # Use lower precision for memory efficiency
        )
        processor = PaliGemmaProcessor.from_pretrained(model_id)
        logger.info("Model loaded successfully")

def clean_memory():
    """Force garbage collection and clear CUDA cache"""
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        # Clear GPU cache
        torch.cuda.empty_cache()
        logger.info(f"Memory allocated after clearing cache: {torch.cuda.memory_allocated()} bytes")
        logger.info("Memory cleaned")

def predict(image):
    """Process a single image"""
    load_model()  # Ensure model is loaded
    
    # Process input
    prompt = "<image> ocr"
    model_inputs = processor(text=prompt, images=image, return_tensors="pt")
    
    # Move to appropriate device
    model_inputs = {k: v.to(model.device) for k, v in model_inputs.items()}
    
    # Generate with memory optimization
    with torch.inference_mode():
        generation = model.generate(**model_inputs, max_new_tokens=200)
        
    # Decode output
    decoded = processor.decode(generation[0], skip_special_tokens=True)
    
    # Clean up intermediates
    del model_inputs, generation
    clean_memory()
    
    return decoded

@app.post("/extract_text")
async def extract_text(background_tasks: BackgroundTasks, file: UploadFile = File(...)):
    """Extract text from a single image"""
    try:
        start_time = time.time()
        image = Image.open(BytesIO(await file.read())).convert("RGB")
        text = predict(image)
        
        # Schedule cleanup after response
        background_tasks.add_task(clean_memory)
        
        logger.info(f"Processing completed in {time.time() - start_time:.2f} seconds")
        return {"extracted_text": text}
    except Exception as e:
        logger.error(f"Error processing image: {str(e)}")
        return {"error": str(e)}

@app.post("/batch_extract_text")
async def batch_extract_text(batch_size:int, background_tasks: BackgroundTasks, files: List[UploadFile] = File(...)):
    """Extract text from multiple images with batching"""
    try:
        start_time = time.time()
        
        # Limit batch size for memory management
        max_batch_size = batch_size  # Adjust based on your GPU memory
        
        # if len(files) > 32:
        #     return {"error": "A maximum of 20 images can be processed at a time."}
        
        load_model()  # Ensure model is loaded
        
        all_results = []
        
        # Process in smaller batches
        for i in range(0, len(files), max_batch_size):
            batch_files = files[i:i+max_batch_size]
            
            # Load images
            images = []
            for file in batch_files:
                image_data = await file.read()
                img = Image.open(BytesIO(image_data)).convert("RGB")
                images.append(img)
            
            # Create batch inputs
            prompts = ["<image> ocr"] * len(images)
            model_inputs = processor(text=prompts, images=images, return_tensors="pt")
            
            # Move to appropriate device
            model_inputs = {k: v.to(model.device) for k, v in model_inputs.items()}
            
            # Generate with memory optimization
            with torch.inference_mode():
                generations = model.generate(**model_inputs, max_new_tokens=200, do_sample=False)
            
            # Decode outputs
            batch_results = [processor.decode(generations[i], skip_special_tokens=True) for i in range(len(images))]
            all_results.extend(batch_results)
            
            # Clean up batch resources
            del model_inputs, generations, images
            clean_memory()
        
        # Schedule cleanup after response
        background_tasks.add_task(clean_memory)
        
        logger.info(f"Batch processing completed in {time.time() - start_time:.2f} seconds")
        return {"extracted_texts": all_results}
    except Exception as e:
        logger.error(f"Error in batch processing: {str(e)}")
        return {"error": str(e)}


# Health check endpoint
@app.get("/health")
async def health_check():
    # Generate a random image (20x40 pixels) with random RGB values
    random_data = np.random.randint(0, 256, (20, 40, 3), dtype=np.uint8)
    
    # Create an image from the random data
    image = Image.fromarray(random_data)
    predict(image)
    clean_memory()
    return {"status": "healthy"}

# if __name__ == "__main__":
#     import uvicorn
    
#     # Start the server with proper worker configuration
#     uvicorn.run(
#         app, 
#         host="0.0.0.0", 
#         port=7860,
#         log_level="info",
#         workers=1  # Multiple workers can cause GPU memory issues
#     )