Spaces:
Running
on
Zero
Running
on
Zero
Create models.py
Browse files
models.py
ADDED
@@ -0,0 +1,362 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Model management for FLUX Prompt Optimizer
|
3 |
+
Handles Florence-2 and Bagel model integration
|
4 |
+
"""
|
5 |
+
|
6 |
+
import logging
|
7 |
+
import requests
|
8 |
+
import spaces
|
9 |
+
import torch
|
10 |
+
from typing import Optional, Dict, Any, Tuple
|
11 |
+
from PIL import Image
|
12 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
13 |
+
|
14 |
+
from config import MODEL_CONFIG, get_device_config
|
15 |
+
from utils import clean_memory, safe_execute
|
16 |
+
|
17 |
+
logger = logging.getLogger(__name__)
|
18 |
+
|
19 |
+
|
20 |
+
class BaseImageAnalyzer:
|
21 |
+
"""Base class for image analysis models"""
|
22 |
+
|
23 |
+
def __init__(self):
|
24 |
+
self.model = None
|
25 |
+
self.processor = None
|
26 |
+
self.device_config = get_device_config()
|
27 |
+
self.is_initialized = False
|
28 |
+
|
29 |
+
def initialize(self) -> bool:
|
30 |
+
"""Initialize the model"""
|
31 |
+
raise NotImplementedError
|
32 |
+
|
33 |
+
def analyze_image(self, image: Image.Image) -> Tuple[str, Dict[str, Any]]:
|
34 |
+
"""Analyze image and return description"""
|
35 |
+
raise NotImplementedError
|
36 |
+
|
37 |
+
def cleanup(self) -> None:
|
38 |
+
"""Clean up model resources"""
|
39 |
+
if self.model is not None:
|
40 |
+
del self.model
|
41 |
+
self.model = None
|
42 |
+
if self.processor is not None:
|
43 |
+
del self.processor
|
44 |
+
self.processor = None
|
45 |
+
clean_memory()
|
46 |
+
|
47 |
+
|
48 |
+
class Florence2Analyzer(BaseImageAnalyzer):
|
49 |
+
"""Florence-2 model for image analysis"""
|
50 |
+
|
51 |
+
def __init__(self):
|
52 |
+
super().__init__()
|
53 |
+
self.config = MODEL_CONFIG["florence2"]
|
54 |
+
|
55 |
+
def initialize(self) -> bool:
|
56 |
+
"""Initialize Florence-2 model"""
|
57 |
+
if self.is_initialized:
|
58 |
+
return True
|
59 |
+
|
60 |
+
try:
|
61 |
+
logger.info("Initializing Florence-2 model...")
|
62 |
+
|
63 |
+
model_id = self.config["model_id"]
|
64 |
+
|
65 |
+
# Load processor
|
66 |
+
self.processor = AutoProcessor.from_pretrained(
|
67 |
+
model_id,
|
68 |
+
trust_remote_code=self.config["trust_remote_code"]
|
69 |
+
)
|
70 |
+
|
71 |
+
# Load model
|
72 |
+
self.model = AutoModelForCausalLM.from_pretrained(
|
73 |
+
model_id,
|
74 |
+
trust_remote_code=self.config["trust_remote_code"],
|
75 |
+
torch_dtype=self.config["torch_dtype"] if self.device_config["use_gpu"] else torch.float32
|
76 |
+
)
|
77 |
+
|
78 |
+
# Move to appropriate device
|
79 |
+
if self.device_config["use_gpu"]:
|
80 |
+
self.model = self.model.to(self.device_config["device"])
|
81 |
+
else:
|
82 |
+
self.model = self.model.to("cpu")
|
83 |
+
|
84 |
+
self.model.eval()
|
85 |
+
self.is_initialized = True
|
86 |
+
|
87 |
+
logger.info(f"Florence-2 initialized on {self.device_config['device']}")
|
88 |
+
return True
|
89 |
+
|
90 |
+
except Exception as e:
|
91 |
+
logger.error(f"Florence-2 initialization failed: {e}")
|
92 |
+
self.cleanup()
|
93 |
+
return False
|
94 |
+
|
95 |
+
@spaces.GPU(duration=60)
|
96 |
+
def _gpu_inference(self, image: Image.Image, task_prompt: str) -> str:
|
97 |
+
"""Run inference on GPU with spaces decorator"""
|
98 |
+
try:
|
99 |
+
# Move model to GPU for inference
|
100 |
+
if self.device_config["use_gpu"]:
|
101 |
+
self.model = self.model.to("cuda")
|
102 |
+
|
103 |
+
# Prepare inputs
|
104 |
+
inputs = self.processor(text=task_prompt, images=image, return_tensors="pt")
|
105 |
+
|
106 |
+
# Move inputs to device
|
107 |
+
device = "cuda" if self.device_config["use_gpu"] else self.device_config["device"]
|
108 |
+
inputs = {k: v.to(device) for k, v in inputs.items()}
|
109 |
+
|
110 |
+
# Generate response
|
111 |
+
with torch.no_grad():
|
112 |
+
if self.device_config["use_gpu"]:
|
113 |
+
with torch.cuda.amp.autocast(dtype=torch.float16):
|
114 |
+
generated_ids = self.model.generate(
|
115 |
+
input_ids=inputs["input_ids"],
|
116 |
+
pixel_values=inputs["pixel_values"],
|
117 |
+
max_new_tokens=self.config["max_new_tokens"],
|
118 |
+
num_beams=3,
|
119 |
+
do_sample=False
|
120 |
+
)
|
121 |
+
else:
|
122 |
+
generated_ids = self.model.generate(
|
123 |
+
input_ids=inputs["input_ids"],
|
124 |
+
pixel_values=inputs["pixel_values"],
|
125 |
+
max_new_tokens=self.config["max_new_tokens"],
|
126 |
+
num_beams=3,
|
127 |
+
do_sample=False
|
128 |
+
)
|
129 |
+
|
130 |
+
# Decode response
|
131 |
+
generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
|
132 |
+
parsed = self.processor.post_process_generation(
|
133 |
+
generated_text,
|
134 |
+
task=task_prompt,
|
135 |
+
image_size=(image.width, image.height)
|
136 |
+
)
|
137 |
+
|
138 |
+
# Extract caption
|
139 |
+
if task_prompt in parsed:
|
140 |
+
return parsed[task_prompt]
|
141 |
+
else:
|
142 |
+
return str(parsed) if parsed else ""
|
143 |
+
|
144 |
+
except Exception as e:
|
145 |
+
logger.error(f"Florence-2 GPU inference failed: {e}")
|
146 |
+
return ""
|
147 |
+
finally:
|
148 |
+
# Move model back to CPU to free GPU memory
|
149 |
+
if self.device_config["use_gpu"]:
|
150 |
+
self.model = self.model.to("cpu")
|
151 |
+
clean_memory()
|
152 |
+
|
153 |
+
def analyze_image(self, image: Image.Image) -> Tuple[str, Dict[str, Any]]:
|
154 |
+
"""Analyze image using Florence-2"""
|
155 |
+
if not self.is_initialized:
|
156 |
+
success = self.initialize()
|
157 |
+
if not success:
|
158 |
+
return "Model initialization failed", {"error": "Florence-2 not available"}
|
159 |
+
|
160 |
+
try:
|
161 |
+
# Define analysis tasks
|
162 |
+
tasks = {
|
163 |
+
"detailed": "<DETAILED_CAPTION>",
|
164 |
+
"more_detailed": "<MORE_DETAILED_CAPTION>",
|
165 |
+
"caption": "<CAPTION>"
|
166 |
+
}
|
167 |
+
|
168 |
+
results = {}
|
169 |
+
|
170 |
+
# Run analysis for each task
|
171 |
+
for task_name, task_prompt in tasks.items():
|
172 |
+
if self.device_config["use_gpu"]:
|
173 |
+
result = self._gpu_inference(image, task_prompt)
|
174 |
+
else:
|
175 |
+
result = self._cpu_inference(image, task_prompt)
|
176 |
+
results[task_name] = result
|
177 |
+
|
178 |
+
# Choose best result
|
179 |
+
if results["more_detailed"]:
|
180 |
+
main_description = results["more_detailed"]
|
181 |
+
elif results["detailed"]:
|
182 |
+
main_description = results["detailed"]
|
183 |
+
else:
|
184 |
+
main_description = results["caption"] or "A photograph"
|
185 |
+
|
186 |
+
# Prepare metadata
|
187 |
+
metadata = {
|
188 |
+
"model": "Florence-2",
|
189 |
+
"device": self.device_config["device"],
|
190 |
+
"all_results": results,
|
191 |
+
"confidence": 0.85 # Florence-2 generally reliable
|
192 |
+
}
|
193 |
+
|
194 |
+
logger.info(f"Florence-2 analysis complete: {len(main_description)} chars")
|
195 |
+
return main_description, metadata
|
196 |
+
|
197 |
+
except Exception as e:
|
198 |
+
logger.error(f"Florence-2 analysis failed: {e}")
|
199 |
+
return "Analysis failed", {"error": str(e)}
|
200 |
+
|
201 |
+
def _cpu_inference(self, image: Image.Image, task_prompt: str) -> str:
|
202 |
+
"""Run inference on CPU"""
|
203 |
+
try:
|
204 |
+
inputs = self.processor(text=task_prompt, images=image, return_tensors="pt")
|
205 |
+
|
206 |
+
with torch.no_grad():
|
207 |
+
generated_ids = self.model.generate(
|
208 |
+
input_ids=inputs["input_ids"],
|
209 |
+
pixel_values=inputs["pixel_values"],
|
210 |
+
max_new_tokens=self.config["max_new_tokens"],
|
211 |
+
num_beams=2, # Reduced for CPU
|
212 |
+
do_sample=False
|
213 |
+
)
|
214 |
+
|
215 |
+
generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
|
216 |
+
parsed = self.processor.post_process_generation(
|
217 |
+
generated_text,
|
218 |
+
task=task_prompt,
|
219 |
+
image_size=(image.width, image.height)
|
220 |
+
)
|
221 |
+
|
222 |
+
if task_prompt in parsed:
|
223 |
+
return parsed[task_prompt]
|
224 |
+
else:
|
225 |
+
return str(parsed) if parsed else ""
|
226 |
+
|
227 |
+
except Exception as e:
|
228 |
+
logger.error(f"Florence-2 CPU inference failed: {e}")
|
229 |
+
return ""
|
230 |
+
|
231 |
+
|
232 |
+
class BagelAnalyzer(BaseImageAnalyzer):
|
233 |
+
"""Bagel-7B model analyzer via API"""
|
234 |
+
|
235 |
+
def __init__(self):
|
236 |
+
super().__init__()
|
237 |
+
self.config = MODEL_CONFIG["bagel"]
|
238 |
+
self.session = requests.Session()
|
239 |
+
|
240 |
+
def initialize(self) -> bool:
|
241 |
+
"""Initialize Bagel analyzer (API-based)"""
|
242 |
+
try:
|
243 |
+
# Test API connectivity
|
244 |
+
test_response = self.session.get(
|
245 |
+
self.config["api_url"],
|
246 |
+
timeout=self.config["timeout"]
|
247 |
+
)
|
248 |
+
|
249 |
+
if test_response.status_code == 200:
|
250 |
+
self.is_initialized = True
|
251 |
+
logger.info("Bagel API connection established")
|
252 |
+
return True
|
253 |
+
else:
|
254 |
+
logger.error(f"Bagel API not accessible: {test_response.status_code}")
|
255 |
+
return False
|
256 |
+
|
257 |
+
except Exception as e:
|
258 |
+
logger.error(f"Bagel initialization failed: {e}")
|
259 |
+
return False
|
260 |
+
|
261 |
+
def analyze_image(self, image: Image.Image) -> Tuple[str, Dict[str, Any]]:
|
262 |
+
"""Analyze image using Bagel-7B API"""
|
263 |
+
if not self.is_initialized:
|
264 |
+
success = self.initialize()
|
265 |
+
if not success:
|
266 |
+
return "Bagel API not available", {"error": "API connection failed"}
|
267 |
+
|
268 |
+
try:
|
269 |
+
# Convert image to base64 or prepare for API call
|
270 |
+
# Note: This is a placeholder - actual implementation would depend on Bagel API format
|
271 |
+
|
272 |
+
# For now, return a placeholder response
|
273 |
+
# In real implementation, you would:
|
274 |
+
# 1. Convert image to required format
|
275 |
+
# 2. Make API call to Bagel endpoint
|
276 |
+
# 3. Parse response
|
277 |
+
|
278 |
+
description = "Detailed image analysis via Bagel-7B (API implementation needed)"
|
279 |
+
metadata = {
|
280 |
+
"model": "Bagel-7B",
|
281 |
+
"method": "API",
|
282 |
+
"confidence": 0.8
|
283 |
+
}
|
284 |
+
|
285 |
+
logger.info("Bagel analysis complete (placeholder)")
|
286 |
+
return description, metadata
|
287 |
+
|
288 |
+
except Exception as e:
|
289 |
+
logger.error(f"Bagel analysis failed: {e}")
|
290 |
+
return "Analysis failed", {"error": str(e)}
|
291 |
+
|
292 |
+
|
293 |
+
class ModelManager:
|
294 |
+
"""Manager for handling multiple analysis models"""
|
295 |
+
|
296 |
+
def __init__(self, preferred_model: str = None):
|
297 |
+
self.preferred_model = preferred_model or MODEL_CONFIG["primary_model"]
|
298 |
+
self.analyzers = {}
|
299 |
+
self.current_analyzer = None
|
300 |
+
|
301 |
+
def get_analyzer(self, model_name: str = None) -> Optional[BaseImageAnalyzer]:
|
302 |
+
"""Get or create analyzer for specified model"""
|
303 |
+
model_name = model_name or self.preferred_model
|
304 |
+
|
305 |
+
if model_name not in self.analyzers:
|
306 |
+
if model_name == "florence2":
|
307 |
+
self.analyzers[model_name] = Florence2Analyzer()
|
308 |
+
elif model_name == "bagel":
|
309 |
+
self.analyzers[model_name] = BagelAnalyzer()
|
310 |
+
else:
|
311 |
+
logger.error(f"Unknown model: {model_name}")
|
312 |
+
return None
|
313 |
+
|
314 |
+
return self.analyzers[model_name]
|
315 |
+
|
316 |
+
def analyze_image(self, image: Image.Image, model_name: str = None) -> Tuple[str, Dict[str, Any]]:
|
317 |
+
"""Analyze image with specified or preferred model"""
|
318 |
+
analyzer = self.get_analyzer(model_name)
|
319 |
+
if analyzer is None:
|
320 |
+
return "No analyzer available", {"error": "Model not found"}
|
321 |
+
|
322 |
+
success, result = safe_execute(analyzer.analyze_image, image)
|
323 |
+
if success:
|
324 |
+
return result
|
325 |
+
else:
|
326 |
+
return "Analysis failed", {"error": result}
|
327 |
+
|
328 |
+
def cleanup_all(self) -> None:
|
329 |
+
"""Clean up all model resources"""
|
330 |
+
for analyzer in self.analyzers.values():
|
331 |
+
analyzer.cleanup()
|
332 |
+
self.analyzers.clear()
|
333 |
+
clean_memory()
|
334 |
+
|
335 |
+
|
336 |
+
# Global model manager instance
|
337 |
+
model_manager = ModelManager()
|
338 |
+
|
339 |
+
|
340 |
+
def analyze_image(image: Image.Image, model_name: str = None) -> Tuple[str, Dict[str, Any]]:
|
341 |
+
"""
|
342 |
+
Convenience function for image analysis
|
343 |
+
|
344 |
+
Args:
|
345 |
+
image: PIL Image to analyze
|
346 |
+
model_name: Optional model name ("florence2" or "bagel")
|
347 |
+
|
348 |
+
Returns:
|
349 |
+
Tuple of (description, metadata)
|
350 |
+
"""
|
351 |
+
return model_manager.analyze_image(image, model_name)
|
352 |
+
|
353 |
+
|
354 |
+
# Export main components
|
355 |
+
__all__ = [
|
356 |
+
"BaseImageAnalyzer",
|
357 |
+
"Florence2Analyzer",
|
358 |
+
"BagelAnalyzer",
|
359 |
+
"ModelManager",
|
360 |
+
"model_manager",
|
361 |
+
"analyze_image"
|
362 |
+
]
|