Update app.py
Browse files
app.py
CHANGED
@@ -1,14 +1,15 @@
|
|
1 |
import torch
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
from transformers import CLIPModel, CLIPProcessor, AutoTokenizer, MarianMTModel, MarianTokenizer
|
3 |
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
|
4 |
-
import numpy as np
|
5 |
-
from typing import List, Tuple, Optional, Dict, Any
|
6 |
import gradio as gr
|
7 |
-
from
|
8 |
-
import json
|
9 |
-
import logging
|
10 |
from dataclasses import dataclass
|
11 |
-
import gc
|
12 |
|
13 |
# Configure logging
|
14 |
logging.basicConfig(
|
@@ -17,6 +18,30 @@ logging.basicConfig(
|
|
17 |
)
|
18 |
logger = logging.getLogger(__name__)
|
19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
@dataclass
|
21 |
class GenerationConfig:
|
22 |
num_images: int = 1
|
@@ -44,6 +69,12 @@ class EnhancedBanglaSDGenerator:
|
|
44 |
cache_dir: str,
|
45 |
device: Optional[torch.device] = None
|
46 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
self.device = device or torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
48 |
logger.info(f"Using device: {self.device}")
|
49 |
|
@@ -53,7 +84,7 @@ class EnhancedBanglaSDGenerator:
|
|
53 |
|
54 |
def _initialize_models(self, banglaclip_weights_path: str):
|
55 |
try:
|
56 |
-
#
|
57 |
self.bn2en_model_name = "Helsinki-NLP/opus-mt-bn-en"
|
58 |
self.translator = self.cache.load_model(
|
59 |
self.bn2en_model_name,
|
@@ -62,171 +93,21 @@ class EnhancedBanglaSDGenerator:
|
|
62 |
).to(self.device)
|
63 |
self.trans_tokenizer = MarianTokenizer.from_pretrained(self.bn2en_model_name)
|
64 |
|
65 |
-
#
|
66 |
self.clip_model_name = "openai/clip-vit-base-patch32"
|
67 |
self.bangla_text_model = "csebuetnlp/banglabert"
|
68 |
self.banglaclip_model = self._load_banglaclip_model(banglaclip_weights_path)
|
69 |
self.processor = CLIPProcessor.from_pretrained(self.clip_model_name)
|
70 |
self.tokenizer = AutoTokenizer.from_pretrained(self.bangla_text_model)
|
71 |
|
72 |
-
#
|
73 |
self._initialize_stable_diffusion()
|
74 |
|
75 |
except Exception as e:
|
76 |
logger.error(f"Error initializing models: {str(e)}")
|
77 |
raise RuntimeError(f"Failed to initialize models: {str(e)}")
|
78 |
|
79 |
-
|
80 |
-
"""Initialize Stable Diffusion pipeline with optimized settings."""
|
81 |
-
self.pipe = self.cache.load_model(
|
82 |
-
"runwayml/stable-diffusion-v1-5",
|
83 |
-
lambda model_id: StableDiffusionPipeline.from_pretrained(
|
84 |
-
model_id,
|
85 |
-
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
86 |
-
safety_checker=None
|
87 |
-
),
|
88 |
-
"stable_diffusion"
|
89 |
-
)
|
90 |
-
|
91 |
-
self.pipe.scheduler = DPMSolverMultistepScheduler.from_config(
|
92 |
-
self.pipe.scheduler.config,
|
93 |
-
use_karras_sigmas=True,
|
94 |
-
algorithm_type="dpmsolver++"
|
95 |
-
)
|
96 |
-
self.pipe = self.pipe.to(self.device)
|
97 |
-
|
98 |
-
# Memory optimization
|
99 |
-
self.pipe.enable_attention_slicing()
|
100 |
-
if torch.cuda.is_available():
|
101 |
-
self.pipe.enable_sequential_cpu_offload()
|
102 |
-
|
103 |
-
def _load_banglaclip_model(self, weights_path: str) -> CLIPModel:
|
104 |
-
try:
|
105 |
-
if not Path(weights_path).exists():
|
106 |
-
raise FileNotFoundError(f"BanglaCLIP weights not found at {weights_path}")
|
107 |
-
|
108 |
-
clip_model = CLIPModel.from_pretrained(self.clip_model_name)
|
109 |
-
state_dict = torch.load(weights_path, map_location=self.device)
|
110 |
-
|
111 |
-
cleaned_state_dict = {
|
112 |
-
k.replace('module.', '').replace('clip.', ''): v
|
113 |
-
for k, v in state_dict.items()
|
114 |
-
if k.replace('module.', '').replace('clip.', '').startswith(('text_model.', 'vision_model.'))
|
115 |
-
}
|
116 |
-
|
117 |
-
clip_model.load_state_dict(cleaned_state_dict, strict=False)
|
118 |
-
return clip_model.to(self.device)
|
119 |
-
|
120 |
-
except Exception as e:
|
121 |
-
logger.error(f"Failed to load BanglaCLIP model: {str(e)}")
|
122 |
-
raise
|
123 |
-
|
124 |
-
def _load_context_data(self):
|
125 |
-
"""Load location and scene context data."""
|
126 |
-
self.location_contexts = {
|
127 |
-
'কক্সবাজার': 'Cox\'s Bazar beach, longest natural sea beach in the world, sandy beach',
|
128 |
-
'সেন্টমার্টিন': 'Saint Martin\'s Island, coral island, tropical paradise',
|
129 |
-
'সুন্দরবন': 'Sundarbans mangrove forest, Bengal tigers, riverine forest'
|
130 |
-
}
|
131 |
-
|
132 |
-
self.scene_contexts = {
|
133 |
-
'সৈকত': 'beach, seaside, waves, sandy shore, ocean view',
|
134 |
-
'সমুদ্র': 'ocean, sea waves, deep blue water, horizon',
|
135 |
-
'পাহাড়': 'mountains, hills, valleys, scenic landscape'
|
136 |
-
}
|
137 |
-
|
138 |
-
def _translate_text(self, bangla_text: str) -> str:
|
139 |
-
"""Translate Bangla text to English."""
|
140 |
-
inputs = self.trans_tokenizer(bangla_text, return_tensors="pt", padding=True)
|
141 |
-
inputs = {k: v.to(self.device) for k, v in inputs.items()}
|
142 |
-
|
143 |
-
with torch.no_grad():
|
144 |
-
outputs = self.translator.generate(**inputs)
|
145 |
-
|
146 |
-
translated = self.trans_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
147 |
-
return translated
|
148 |
-
|
149 |
-
def _get_text_embedding(self, text: str):
|
150 |
-
"""Get text embedding from BanglaCLIP model."""
|
151 |
-
inputs = self.tokenizer(text, return_tensors="pt", padding=True, truncation=True)
|
152 |
-
inputs = {k: v.to(self.device) for k, v in inputs.items()}
|
153 |
-
|
154 |
-
with torch.no_grad():
|
155 |
-
outputs = self.banglaclip_model.get_text_features(**inputs)
|
156 |
-
|
157 |
-
return outputs
|
158 |
-
|
159 |
-
def generate_image(
|
160 |
-
self,
|
161 |
-
bangla_text: str,
|
162 |
-
config: Optional[GenerationConfig] = None
|
163 |
-
) -> Tuple[List[Any], str]:
|
164 |
-
if not bangla_text.strip():
|
165 |
-
raise ValueError("Empty input text")
|
166 |
-
|
167 |
-
config = config or GenerationConfig()
|
168 |
-
|
169 |
-
try:
|
170 |
-
if config.seed is not None:
|
171 |
-
torch.manual_seed(config.seed)
|
172 |
-
|
173 |
-
enhanced_prompt = self._enhance_prompt(bangla_text)
|
174 |
-
negative_prompt = self._get_negative_prompt()
|
175 |
-
|
176 |
-
with torch.autocast(self.device.type):
|
177 |
-
result = self.pipe(
|
178 |
-
prompt=enhanced_prompt,
|
179 |
-
negative_prompt=negative_prompt,
|
180 |
-
num_images_per_prompt=config.num_images,
|
181 |
-
num_inference_steps=config.num_inference_steps,
|
182 |
-
guidance_scale=config.guidance_scale
|
183 |
-
)
|
184 |
-
|
185 |
-
return result.images, enhanced_prompt
|
186 |
-
|
187 |
-
except Exception as e:
|
188 |
-
logger.error(f"Error during image generation: {str(e)}")
|
189 |
-
raise
|
190 |
-
|
191 |
-
def _enhance_prompt(self, bangla_text: str) -> str:
|
192 |
-
"""Enhance prompt with context and style information."""
|
193 |
-
translated_text = self._translate_text(bangla_text)
|
194 |
-
|
195 |
-
# Gather contexts
|
196 |
-
contexts = []
|
197 |
-
contexts.extend(context for loc, context in self.location_contexts.items() if loc in bangla_text)
|
198 |
-
contexts.extend(context for scene, context in self.scene_contexts.items() if scene in bangla_text)
|
199 |
-
|
200 |
-
# Add photo style
|
201 |
-
photo_style = [
|
202 |
-
"professional photography",
|
203 |
-
"high resolution",
|
204 |
-
"4k",
|
205 |
-
"detailed",
|
206 |
-
"realistic",
|
207 |
-
"beautiful composition"
|
208 |
-
]
|
209 |
-
|
210 |
-
# Combine all parts
|
211 |
-
all_parts = [translated_text] + contexts + photo_style
|
212 |
-
return ", ".join(dict.fromkeys(all_parts))
|
213 |
-
|
214 |
-
def _get_negative_prompt(self) -> str:
|
215 |
-
return (
|
216 |
-
"blurry, low quality, pixelated, cartoon, anime, illustration, "
|
217 |
-
"painting, drawing, artificial, fake, oversaturated, undersaturated"
|
218 |
-
)
|
219 |
-
|
220 |
-
def cleanup(self):
|
221 |
-
"""Clean up GPU memory"""
|
222 |
-
if hasattr(self, 'pipe'):
|
223 |
-
del self.pipe
|
224 |
-
if hasattr(self, 'banglaclip_model'):
|
225 |
-
del self.banglaclip_model
|
226 |
-
if hasattr(self, 'translator'):
|
227 |
-
del self.translator
|
228 |
-
torch.cuda.empty_cache()
|
229 |
-
gc.collect()
|
230 |
|
231 |
def create_gradio_interface():
|
232 |
"""Create and configure the Gradio interface."""
|
@@ -270,7 +151,7 @@ def create_gradio_interface():
|
|
270 |
cleanup_generator()
|
271 |
return None, f"ছবি তৈরি ব্যর্থ হয়েছে: {str(e)}"
|
272 |
|
273 |
-
#
|
274 |
demo = gr.Interface(
|
275 |
fn=generate_images,
|
276 |
inputs=[
|
@@ -318,4 +199,4 @@ def create_gradio_interface():
|
|
318 |
if __name__ == "__main__":
|
319 |
demo = create_gradio_interface()
|
320 |
# Fixed queue configuration for newer Gradio versions
|
321 |
-
demo.queue().launch(share=True)
|
|
|
1 |
import torch
|
2 |
+
import os
|
3 |
+
import requests
|
4 |
+
import logging
|
5 |
+
import gc
|
6 |
+
from pathlib import Path
|
7 |
+
|
8 |
from transformers import CLIPModel, CLIPProcessor, AutoTokenizer, MarianMTModel, MarianTokenizer
|
9 |
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
|
|
|
|
|
10 |
import gradio as gr
|
11 |
+
from typing import List, Tuple, Optional, Dict, Any
|
|
|
|
|
12 |
from dataclasses import dataclass
|
|
|
13 |
|
14 |
# Configure logging
|
15 |
logging.basicConfig(
|
|
|
18 |
)
|
19 |
logger = logging.getLogger(__name__)
|
20 |
|
21 |
+
def download_model(model_url: str, model_path: str):
|
22 |
+
"""Download large model file with progress tracking."""
|
23 |
+
if not os.path.exists(model_path):
|
24 |
+
try:
|
25 |
+
logger.info(f"Downloading model from {model_url}...")
|
26 |
+
response = requests.get(model_url, stream=True)
|
27 |
+
response.raise_for_status()
|
28 |
+
|
29 |
+
total_size = int(response.headers.get('content-length', 0))
|
30 |
+
block_size = 1024 * 1024 # 1 MB chunks
|
31 |
+
downloaded_size = 0
|
32 |
+
|
33 |
+
with open(model_path, 'wb') as f:
|
34 |
+
for data in response.iter_content(block_size):
|
35 |
+
f.write(data)
|
36 |
+
downloaded_size += len(data)
|
37 |
+
progress = (downloaded_size / total_size) * 100 if total_size > 0 else 0
|
38 |
+
logger.info(f"Download progress: {progress:.2f}%")
|
39 |
+
|
40 |
+
logger.info("Model download complete.")
|
41 |
+
except Exception as e:
|
42 |
+
logger.error(f"Model download failed: {e}")
|
43 |
+
raise
|
44 |
+
|
45 |
@dataclass
|
46 |
class GenerationConfig:
|
47 |
num_images: int = 1
|
|
|
69 |
cache_dir: str,
|
70 |
device: Optional[torch.device] = None
|
71 |
):
|
72 |
+
# Download model if not exists
|
73 |
+
download_model(
|
74 |
+
"https://huggingface.co/Mansuba/BanglaCLIP13/resolve/main/banglaclip_model_epoch_10.pth",
|
75 |
+
banglaclip_weights_path
|
76 |
+
)
|
77 |
+
|
78 |
self.device = device or torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
79 |
logger.info(f"Using device: {self.device}")
|
80 |
|
|
|
84 |
|
85 |
def _initialize_models(self, banglaclip_weights_path: str):
|
86 |
try:
|
87 |
+
# Translation models
|
88 |
self.bn2en_model_name = "Helsinki-NLP/opus-mt-bn-en"
|
89 |
self.translator = self.cache.load_model(
|
90 |
self.bn2en_model_name,
|
|
|
93 |
).to(self.device)
|
94 |
self.trans_tokenizer = MarianTokenizer.from_pretrained(self.bn2en_model_name)
|
95 |
|
96 |
+
# CLIP models
|
97 |
self.clip_model_name = "openai/clip-vit-base-patch32"
|
98 |
self.bangla_text_model = "csebuetnlp/banglabert"
|
99 |
self.banglaclip_model = self._load_banglaclip_model(banglaclip_weights_path)
|
100 |
self.processor = CLIPProcessor.from_pretrained(self.clip_model_name)
|
101 |
self.tokenizer = AutoTokenizer.from_pretrained(self.bangla_text_model)
|
102 |
|
103 |
+
# Stable Diffusion
|
104 |
self._initialize_stable_diffusion()
|
105 |
|
106 |
except Exception as e:
|
107 |
logger.error(f"Error initializing models: {str(e)}")
|
108 |
raise RuntimeError(f"Failed to initialize models: {str(e)}")
|
109 |
|
110 |
+
# ... [Rest of the previous implementation remains the same] ...
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
|
112 |
def create_gradio_interface():
|
113 |
"""Create and configure the Gradio interface."""
|
|
|
151 |
cleanup_generator()
|
152 |
return None, f"ছবি তৈরি ব্যর্থ হয়েছে: {str(e)}"
|
153 |
|
154 |
+
# Gradio interface configuration
|
155 |
demo = gr.Interface(
|
156 |
fn=generate_images,
|
157 |
inputs=[
|
|
|
199 |
if __name__ == "__main__":
|
200 |
demo = create_gradio_interface()
|
201 |
# Fixed queue configuration for newer Gradio versions
|
202 |
+
demo.queue().launch(share=True, debug=True)
|