VOIDER commited on
Commit
e84a5b4
·
verified ·
1 Parent(s): f56b01d

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +432 -884
app.py CHANGED
@@ -1,73 +1,76 @@
1
  import os
2
- import shutil
3
  import tempfile
4
- import asyncio
5
- from io import BytesIO, StringIO
6
- import csv
7
- from pathlib import Path
8
- import logging
9
 
10
  import cv2
11
  import numpy as np
 
12
  import torch
13
  import onnxruntime as rt
14
  from PIL import Image
15
  import gradio as gr
16
- from transformers import pipeline, AutoProcessor, AutoModelForImageClassification
17
  from huggingface_hub import hf_hub_download
18
 
19
- # Configure basic logging
20
- logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
21
- logger = logging.getLogger(__name__)
22
-
23
- # --- Dependency: aesthetic_predictor_v2_5.py ---
24
- # This file should exist in the same directory or be in PYTHONPATH.
25
- # For demonstration, a stub is provided. Replace with actual implementation.
26
- # aesthetic_predictor_v2_5.py STUB START
27
- # (Normally this would be in its own file: aesthetic_predictor_v2_5.py)
28
- def convert_v2_5_from_siglip(repo_id="unum-cloud/siglip-base-patch16-224-aesthetic-v2.5", low_cpu_mem_usage=True, trust_remote_code=True):
29
- logger.info(f"Loading model and preprocessor from Hugging Face Hub: {repo_id}")
30
- try:
31
- # Attempt to load actual models if available and network permits
32
- processor = AutoProcessor.from_pretrained(repo_id, low_cpu_mem_usage=low_cpu_mem_usage, trust_remote_code=trust_remote_code)
33
- model = AutoModelForImageClassification.from_pretrained(repo_id, low_cpu_mem_usage=low_cpu_mem_usage, trust_remote_code=trust_remote_code)
34
- logger.info("Successfully loaded model and preprocessor from Hugging Face Hub.")
35
- except Exception as e:
36
- logger.warning(f"Failed to load from {repo_id} due to: {e}. Using fallback mock objects.")
37
- # Fallback to simpler mock objects if HF download fails or for offline use
38
- class MockProcessor:
39
- def __call__(self, images, return_tensors="pt"):
40
- if isinstance(images, list):
41
- num_images = len(images)
42
- return {"pixel_values": torch.randn(num_images, 3, 224, 224)}
43
- else:
44
- return {"pixel_values": torch.randn(1, 3, 224, 224)}
45
- class MockModel:
46
- def __init__(self): self._parameters = {"dummy": torch.nn.Parameter(torch.empty(0))}
47
- def __call__(self, pixel_values):
48
- bs = pixel_values.shape[0]
49
  class Output:
50
- def __init__(self, logits_val): self.logits = logits_val
51
- return Output(logits_val=torch.rand(bs, 1) * 10) # Simulate scores 0-10
52
- def to(self, *args, **kwargs): return self
53
- def cuda(self, *args, **kwargs): return self
54
- def bfloat16(self, *args, **kwargs): return self
55
- processor = MockProcessor()
56
- model = MockModel()
57
- logger.info("Using fallback mock model and preprocessor for Aesthetic Predictor V2.5.")
58
- return model, processor
59
- # aesthetic_predictor_v2_5.py STUB END
60
-
61
-
62
- #####################################
63
- # Model Definitions #
64
- #####################################
 
 
 
 
 
 
 
 
 
 
 
 
 
 
65
 
66
  class MLP(torch.nn.Module):
 
67
  def __init__(self, input_size: int, batch_norm: bool = True):
68
  super().__init__()
69
  self.input_size = input_size
70
- layers = [
71
  torch.nn.Linear(self.input_size, 2048), torch.nn.ReLU(),
72
  torch.nn.BatchNorm1d(2048) if batch_norm else torch.nn.Identity(), torch.nn.Dropout(0.3),
73
  torch.nn.Linear(2048, 512), torch.nn.ReLU(),
@@ -78,916 +81,461 @@ class MLP(torch.nn.Module):
78
  torch.nn.BatchNorm1d(128) if batch_norm else torch.nn.Identity(), torch.nn.Dropout(0.1),
79
  torch.nn.Linear(128, 32), torch.nn.ReLU(),
80
  torch.nn.Linear(32, 1)
81
- ]
82
- self.layers = torch.nn.Sequential(*layers)
83
-
84
- def forward(self, x: torch.Tensor) -> torch.Tensor:
85
- return self.layers(x)
86
-
87
 
88
- class WaifuScorer:
89
- def __init__(self, model_path: str = None, device: str = 'cuda', cache_dir: str = None, verbose: bool = False):
90
- self.verbose = verbose
 
 
91
  self.device = device
92
- self.dtype = torch.float32
93
- self.available = False
94
- self.clip_model = None
95
- self.preprocess = None
96
- self.mlp = None
97
-
98
- try:
99
- import clip # Dynamically import clip
100
- if model_path is None:
101
- model_path = "Eugeoter/waifu-scorer-v3/model.pth"
102
- if self.verbose: logger.info(f"WaifuScorer model path not provided. Using default: {model_path}")
103
 
104
- if not Path(model_path).is_file():
105
- try:
106
- # Assuming model_path like "user/repo/file.pth" for hf_hub_download
107
- parts = model_path.split("/")
108
- if len(parts) >= 3:
109
- repo_id_parts = parts[:-1]
110
- filename = parts[-1]
111
- repo_id_str = "/".join(repo_id_parts)
112
- model_path_resolved = hf_hub_download(repo_id=repo_id_str, filename=filename, cache_dir=cache_dir)
113
- else: # try as repo_id and assume model.pth or common name
114
- model_path_resolved = hf_hub_download(repo_id=model_path, filename="model.pth", cache_dir=cache_dir) # fallback filename
115
- except Exception as e:
116
- logger.error(f"Failed to download WaifuScorer model from HF Hub ({model_path}): {e}")
117
- # Try a more specific default if the generic one failed
118
- logger.info("Attempting to download specific WaifuScorer model Eugeoter/waifu-scorer-v3/model.pth")
119
- model_path_resolved = hf_hub_download("Eugeoter/waifu-scorer-v3", "model.pth", cache_dir=cache_dir)
120
- model_path = model_path_resolved
121
 
 
 
 
 
 
 
 
122
 
123
- if self.verbose: logger.info(f"Loading WaifuScorer model from: {model_path}")
 
 
 
 
 
 
 
 
124
 
125
- self.mlp = MLP(input_size=768)
126
- if str(model_path).endswith(".safetensors"):
127
  from safetensors.torch import load_file
128
- state_dict = load_file(model_path, device=device)
129
  else:
130
- state_dict = torch.load(model_path, map_location=device)
131
-
132
- # Adjust keys if necessary (e.g. if saved from DataParallel)
133
- if any(key.startswith("module.") for key in state_dict.keys()):
134
- state_dict = {k.replace("module.", ""): v for k, v in state_dict.items()}
135
-
136
  self.mlp.load_state_dict(state_dict)
137
- self.mlp.to(device=self.device, dtype=self.dtype)
138
- self.mlp.eval()
139
 
140
- self.clip_model, self.preprocess = clip.load("ViT-L/14", device=self.device)
141
- self.available = True
142
- logger.info("WaifuScorer initialized successfully.")
143
  except ImportError:
144
- logger.error("OpenAI CLIP library not found. WaifuScorer will be unavailable. Please install with 'pip install openai-clip'")
145
  except Exception as e:
146
- logger.error(f"Unable to initialize WaifuScorer: {e}")
147
 
148
  @torch.no_grad()
149
- def __call__(self, images: list[Image.Image]) -> list[float | None]:
150
- if not self.available:
151
- return [None] * len(images)
152
 
153
- if not images:
154
- return []
155
-
156
  original_n = len(images)
157
- # Handle single image case for CLIP if it has issues with batch_size=1 (some versions might)
158
- processed_images = images if len(images) > 1 else images * 2
159
 
160
  try:
161
- image_tensors = [self.preprocess(img).unsqueeze(0) for img in processed_images]
162
- image_batch = torch.cat(image_tensors).to(self.device)
163
- image_features = self.clip_model.encode_image(image_batch)
164
-
165
  norm = image_features.norm(p=2, dim=-1, keepdim=True)
166
- norm = torch.where(norm == 0, torch.tensor(1.0, device=norm.device, dtype=norm.dtype), norm) # Avoid division by zero
167
- im_emb = (image_features / norm).to(device=self.device, dtype=self.dtype)
168
 
169
  predictions = self.mlp(im_emb)
170
- scores = predictions.clamp(0, 10).cpu().numpy().reshape(-1).tolist()
171
  return scores[:original_n]
172
  except Exception as e:
173
- logger.error(f"Error during WaifuScorer prediction: {e}")
174
  return [None] * original_n
175
 
176
-
177
- class AestheticPredictorV2_5_Wrapper:
178
- def __init__(self, device: str):
179
- self.device = device
180
- self.model, self.preprocessor = convert_v2_5_from_siglip(
181
- low_cpu_mem_usage=True, trust_remote_code=True
182
- )
183
- if self.device == 'cuda' and torch.cuda.is_available():
184
- self.model = self.model.to(torch.bfloat16).cuda()
185
- logger.info("Aesthetic Predictor V2.5 Wrapper initialized.")
 
 
186
 
187
  @torch.no_grad()
188
- def inference(self, images: list[Image.Image]) -> list[float | None]:
189
- if not images:
190
- return []
191
  try:
192
- images_rgb = [img.convert("RGB") for img in images]
193
- pixel_values = self.preprocessor(images=images_rgb, return_tensors="pt").pixel_values
194
- if self.device == 'cuda' and torch.cuda.is_available():
195
- pixel_values = pixel_values.to(torch.bfloat16).cuda()
 
 
 
 
196
 
197
- scores_tensor = self.model(pixel_values).logits.squeeze().float().cpu().numpy()
198
- if scores_tensor.ndim == 0: # Single image result
199
- scores = [scores_tensor.item()]
200
- else:
201
- scores = scores_tensor.tolist()
202
- return [round(max(0.0, min(s, 10.0)), 4) for s in scores] # Clip and round
203
  except Exception as e:
204
- logger.error(f"Error during Aesthetic Predictor V2.5 inference: {e}")
205
  return [None] * len(images)
206
 
207
- def load_anime_aesthetic_onnx_model(cache_dir: str = None) -> rt.InferenceSession | None:
208
- try:
209
- model_path = hf_hub_download(repo_id="skytnt/anime-aesthetic", filename="model.onnx", cache_dir=cache_dir)
210
- providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if torch.cuda.is_available() else ['CPUExecutionProvider']
211
- session = rt.InferenceSession(model_path, providers=providers)
212
- logger.info(f"Anime Aesthetic ONNX model loaded with providers: {session.get_providers()}")
213
- return session
214
- except Exception as e:
215
- logger.error(f"Failed to load Anime Aesthetic ONNX model: {e}")
216
- return None
217
-
218
- def preprocess_anime_aesthetic_batch(images_pil: list[Image.Image], target_size: int = 768) -> np.ndarray | None:
219
- if not images_pil:
220
- return None
221
- batch_canvases = []
222
- try:
223
- for img_pil in images_pil:
224
- img_np = np.array(img_pil.convert("RGB")).astype(np.float32) / 255.0
225
- h, w = img_np.shape[:2]
226
- if h > w:
227
- new_h, new_w = target_size, int(target_size * w / h)
228
- else:
229
- new_h, new_w = int(target_size * h / w), target_size
230
-
231
- resized = cv2.resize(img_np, (new_w, new_h), interpolation=cv2.INTER_AREA)
232
- canvas = np.zeros((target_size, target_size, 3), dtype=np.float32)
233
- pad_h = (target_size - new_h) // 2
234
- pad_w = (target_size - new_w) // 2
235
- canvas[pad_h:pad_h+new_h, pad_w:pad_w+new_w] = resized
236
- batch_canvases.append(canvas)
237
-
238
- input_tensor_batch = np.array(batch_canvases, dtype=np.float32) # (N, H, W, C)
239
- input_tensor_batch = np.transpose(input_tensor_batch, (0, 3, 1, 2)) # (N, C, H, W)
240
- return input_tensor_batch
241
- except Exception as e:
242
- logger.error(f"Error during Anime Aesthetic preprocessing: {e}")
243
- return None
244
-
245
- #####################################
246
- # Image Evaluation Tool #
247
- #####################################
248
-
249
- class ModelManager:
250
- def __init__(self, cache_dir: str = None):
251
- self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
252
- logger.info(f"Using device: {self.device}")
253
- self.cache_dir = cache_dir
254
- self.models = {}
255
- self.model_configs = {}
256
- self._load_all_models()
257
-
258
- self.processing_queue: asyncio.Queue = asyncio.Queue()
259
- self.worker_task = None
260
- self._temp_files_to_clean = [] # For CSV files
261
-
262
- def _load_all_models(self):
263
- logger.info("Loading Aesthetic Shadow model...")
264
- try:
265
- self.models["aesthetic_shadow"] = pipeline("image-classification", model="NeoChen1024/aesthetic-shadow-v2-backup", device=0 if self.device == 'cuda' else -1)
266
- self.model_configs["aesthetic_shadow"] = {"name": "Aesthetic Shadow", "process_func": self._process_aesthetic_shadow}
267
- logger.info("Aesthetic Shadow model loaded.")
268
- except Exception as e:
269
- logger.error(f"Failed to load Aesthetic Shadow model: {e}")
270
-
271
- logger.info("Loading Waifu Scorer model...")
272
- try:
273
- ws = WaifuScorer(device=self.device, cache_dir=self.cache_dir, verbose=True)
274
- if ws.available:
275
- self.models["waifu_scorer"] = ws
276
- self.model_configs["waifu_scorer"] = {"name": "Waifu Scorer", "process_func": self._process_waifu_scorer}
277
- logger.info("Waifu Scorer model loaded.")
278
- else:
279
- logger.warning("Waifu Scorer model is not available.")
280
- except Exception as e:
281
- logger.error(f"Failed to load Waifu Scorer model: {e}")
282
-
283
- logger.info("Loading Aesthetic Predictor V2.5...")
284
- try:
285
- ap_v25 = AestheticPredictorV2_5_Wrapper(device=self.device)
286
- self.models["aesthetic_predictor_v2_5"] = ap_v25
287
- self.model_configs["aesthetic_predictor_v2_5"] = {"name": "Aesthetic V2.5", "process_func": self._process_aesthetic_predictor_v2_5}
288
- logger.info("Aesthetic Predictor V2.5 loaded.")
289
- except Exception as e:
290
- logger.error(f"Failed to load Aesthetic Predictor V2.5: {e}")
291
-
292
- logger.info("Loading Anime Aesthetic model...")
293
  try:
294
- aa_model = load_anime_aesthetic_onnx_model(cache_dir=self.cache_dir)
295
- if aa_model:
296
- self.models["anime_aesthetic"] = aa_model
297
- self.model_configs["anime_aesthetic"] = {"name": "Anime Score", "process_func": self._process_anime_aesthetic}
298
- logger.info("Anime Aesthetic model loaded.")
299
- else:
300
- logger.warning("Anime Aesthetic ONNX model failed to load and will be unavailable.")
301
  except Exception as e:
302
- logger.error(f"Failed to load Anime Aesthetic model: {e}")
303
-
304
- logger.info(f"Available models for processing: {list(self.model_configs.keys())}")
305
-
306
-
307
- async def start_worker_if_not_running(self):
308
- if self.worker_task is None or self.worker_task.done():
309
- self.worker_task = asyncio.create_task(self._worker())
310
- logger.info("Async worker started.")
311
-
312
- async def _worker(self):
313
- while True:
314
- request = await self.processing_queue.get()
315
- if request is None: # Shutdown signal
316
- self.processing_queue.task_done()
317
- logger.info("Async worker received shutdown signal.")
318
- break
319
-
320
- future = request.get('future')
321
- try:
322
- if request['type'] == 'run_evaluation_generator':
323
- # The generator itself is created here and returned via future
324
- # The Gradio callback will iterate over it
325
- gen = self.run_evaluation_generator(**request['params'])
326
- future.set_result(gen)
327
- else:
328
- logger.warning(f"Unknown request type in worker: {request.get('type')}")
329
- if future: future.set_exception(ValueError("Unknown request type"))
330
- except Exception as e:
331
- logger.error(f"Error in worker processing request: {e}", exc_info=True)
332
- if future: future.set_exception(e)
333
- finally:
334
- self.processing_queue.task_done()
335
-
336
- async def submit_evaluation_request(self, file_paths, auto_batch, manual_batch_size, selected_model_keys):
337
- await self.start_worker_if_not_running()
338
- future = asyncio.Future()
339
- request_item = {
340
- 'type': 'run_evaluation_generator',
341
- 'params': {
342
- 'file_paths': file_paths,
343
- 'auto_batch': auto_batch,
344
- 'manual_batch_size': manual_batch_size,
345
- 'selected_model_keys': selected_model_keys,
346
- },
347
- 'future': future
348
- }
349
- await self.processing_queue.put(request_item)
350
- return await future # Future resolves to the async generator
351
-
352
- def auto_tune_batch_size(self, images: list[Image.Image], selected_model_keys: list[str]) -> int:
353
- if not images or not selected_model_keys:
354
- return 1
355
 
356
- batch_size = 1
357
- max_possible_batch = len(images)
358
- test_image_pil = [images[0].copy()] # A list containing one PIL image, copy to avoid issues with transforms
359
-
360
- logger.info(f"Auto-tuning batch size with selected models: {selected_model_keys}")
361
 
362
- optimal_batch_size = 1
363
- while batch_size <= max_possible_batch:
364
- current_test_batch = test_image_pil * batch_size
365
- try:
366
- logger.debug(f"Testing batch size: {batch_size}")
367
- if "aesthetic_shadow" in selected_model_keys and "aesthetic_shadow" in self.models:
368
- _ = self.models["aesthetic_shadow"](current_test_batch, batch_size=batch_size)
369
- if "waifu_scorer" in selected_model_keys and "waifu_scorer" in self.models:
370
- _ = self.models["waifu_scorer"](current_test_batch)
371
- if "aesthetic_predictor_v2_5" in selected_model_keys and "aesthetic_predictor_v2_5" in self.models:
372
- _ = self.models["aesthetic_predictor_v2_5"].inference(current_test_batch)
373
- if "anime_aesthetic" in selected_model_keys and "anime_aesthetic" in self.models:
374
- processed_input = preprocess_anime_aesthetic_batch(current_test_batch)
375
- if processed_input is None: raise ValueError("Anime aesthetic preprocessing failed for test batch")
376
- _ = self.models["anime_aesthetic"].run(None, {"img": processed_input})
377
-
378
- optimal_batch_size = batch_size # This batch size worked
379
- if batch_size * 2 > max_possible_batch : # If next step exceeds max, current is best fit
380
- if max_possible_batch > batch_size: # Check if we can exactly fit max_possible_batch
381
- # Test max_possible_batch one last time if it's > current batch_size and < batch_size*2
382
- pass # Current optimal_batch_size is good, or we can check max_possible_batch specifically
383
- break
384
- batch_size *= 2
385
-
386
- except Exception as e: # Typically torch.cuda.OutOfMemoryError or similar
387
- logger.warning(f"Auto-tune failed at batch size {batch_size} for at least one model: {e}")
388
- break # Current optimal_batch_size is the largest that worked before this failure
389
-
390
- # Cap the batch size for very large numbers of images / powerful GPUs
391
- final_optimal_batch = min(optimal_batch_size, max_possible_batch, 64)
392
- logger.info(f"Optimal batch size determined: {final_optimal_batch}")
393
- return max(1, final_optimal_batch)
394
-
395
 
396
- async def run_evaluation_generator(self, file_paths: list[str], auto_batch: bool,
397
- manual_batch_size: int, selected_model_keys: list[str]):
398
-
399
- log_messages = []
400
- def _log(msg):
401
- log_messages.append(msg)
402
- logger.info(msg)
403
-
404
- _log("Starting image evaluation...")
405
- yield {"type": "log_update", "messages": log_messages[-20:]} # Show last 20 logs
406
- yield {"type": "progress", "value": 0.0, "desc": "Initiating..."}
407
-
408
- images_pil = []
409
- file_names = []
410
- for f_path_str in file_paths:
411
  try:
412
- p = Path(f_path_str)
413
- img = Image.open(p).convert("RGB")
414
- images_pil.append(img)
415
- file_names.append(p.name)
416
- _log(f"Loaded image: {p.name}")
417
  except Exception as e:
418
- _log(f"Error opening {f_path_str}: {e}")
419
-
420
- yield {"type": "log_update", "messages": log_messages[-20:]}
421
-
422
- if not images_pil:
423
- _log("No valid images loaded. Aborting.")
424
- yield {"type": "log_update", "messages": log_messages[-20:]}
425
- yield {"type": "progress", "value": 1.0, "desc": "No images loaded"}
426
- yield {"type": "final_results_state", "data": []} # ensure state is empty
427
- return
428
-
429
- actual_batch_size = 1
430
- if auto_batch:
431
- _log("Auto-tuning batch size...")
432
- yield {"type": "log_update", "messages": log_messages[-20:]}
433
- yield {"type": "progress", "value": 0.05, "desc": "Auto-tuning batch size..."}
434
- actual_batch_size = self.auto_tune_batch_size(images_pil, selected_model_keys)
435
- _log(f"Auto-detected batch size: {actual_batch_size}")
436
- else:
437
- actual_batch_size = int(manual_batch_size) if manual_batch_size > 0 else 1
438
- _log(f"Using manual batch size: {actual_batch_size}")
439
-
440
- yield {"type": "batch_size_update", "value": actual_batch_size}
441
- yield {"type": "log_update", "messages": log_messages[-20:]}
442
-
443
- all_results_for_state = [] # Full data for gr.State
444
- dataframe_rows_so_far = [] # Data for gr.DataFrame (PIL images, strings, numbers)
445
-
446
- total_images = len(images_pil)
447
- processed_count = 0
448
-
449
- for i in range(0, total_images, actual_batch_size):
450
- batch_images_pil = images_pil[i:i+actual_batch_size]
451
- batch_file_names = file_names[i:i+actual_batch_size]
452
- num_in_batch = len(batch_images_pil)
453
- _log(f"Processing batch {i//actual_batch_size + 1}/{ (total_images + actual_batch_size -1) // actual_batch_size }: images {i+1} to {i+num_in_batch}")
454
- yield {"type": "log_update", "messages": log_messages[-20:]}
455
-
456
- batch_model_scores = {key: [None] * num_in_batch for key in self.model_configs.keys()}
457
-
458
- for model_key in selected_model_keys:
459
- if model_key in self.models and model_key in self.model_configs:
460
- _log(f" Running {self.model_configs[model_key]['name']} for batch...")
461
- yield {"type": "log_update", "messages": log_messages[-20:]}
462
- try:
463
- scores = await self.model_configs[model_key]['process_func'](batch_images_pil)
464
- batch_model_scores[model_key] = scores
465
- _log(f" {self.model_configs[model_key]['name']} scores: {scores}")
466
- except Exception as e:
467
- _log(f" Error processing batch with {self.model_configs[model_key]['name']}: {e}")
468
- batch_model_scores[model_key] = [None] * num_in_batch # Ensure it's list of Nones
469
- yield {"type": "log_update", "messages": log_messages[-20:]}
470
-
471
- # Assemble results for this batch
472
- current_batch_df_rows = []
473
- for j in range(num_in_batch):
474
- result_item_state = {'file_name': batch_file_names[j]} # For gr.State
475
-
476
- # For DataFrame: [PIL.Image, filename, score1, score2, ..., final_score]
477
- thumbnail = batch_images_pil[j].copy()
478
- thumbnail.thumbnail((150, 150)) # Create thumbnail
479
- result_item_df_row = [thumbnail, batch_file_names[j]]
480
-
481
-
482
- current_image_scores = []
483
- for model_key in self.model_configs.keys(): # Iterate in defined order for consistency
484
- score = batch_model_scores[model_key][j]
485
- result_item_state[model_key] = score # For gr.State
486
- if model_key in selected_model_keys: # Only add to DF if selected
487
- result_item_df_row.append(f"{score:.4f}" if isinstance(score, (float, int)) else "N/A")
488
- if isinstance(score, (float, int)) and model_key in selected_model_keys:
489
- current_image_scores.append(score)
490
-
491
- final_score = None
492
- if current_image_scores:
493
- final_score_val = float(np.mean([s for s in current_image_scores if s is not None]))
494
- final_score = float(np.clip(final_score_val, 0.0, 10.0))
495
-
496
- result_item_state['final_score'] = final_score
497
- result_item_df_row.append(f"{final_score:.4f}" if final_score is not None else "N/A")
498
-
499
- all_results_for_state.append(result_item_state)
500
- current_batch_df_rows.append(result_item_df_row)
501
-
502
- dataframe_rows_so_far.extend(current_batch_df_rows)
503
-
504
- processed_count += num_in_batch
505
- progress_value = processed_count / total_images
506
- yield {"type": "partial_results_df_rows", "data": dataframe_rows_so_far, "selected_model_keys": selected_model_keys}
507
- yield {"type": "progress", "value": progress_value, "desc": f"Processed {processed_count}/{total_images}"}
508
-
509
- _log("All images processed.")
510
- yield {"type": "log_update", "messages": log_messages[-20:]}
511
- yield {"type": "progress", "value": 1.0, "desc": "Completed!"}
512
- yield {"type": "final_results_state", "data": all_results_for_state}
513
 
 
 
 
 
 
 
 
 
514
 
515
- async def _process_aesthetic_shadow(self, batch_images: list[Image.Image]) -> list[float | None]:
516
- model = self.models.get("aesthetic_shadow")
517
- if not model: return [None] * len(batch_images)
518
- results = model(batch_images, batch_size=len(batch_images)) # Assuming pipeline can take batch_size hint
519
  scores = []
520
- for res_group in results: # Results might be List[List[Dict]] or List[Dict]
521
- # Handle both single image and batch results from pipeline
522
- current_res_list = res_group if isinstance(res_group, list) else [res_group]
523
- try:
524
- hq_score_item = next(p for p in current_res_list if p['label'] == 'hq')
525
- score = float(np.clip(hq_score_item['score'] * 10.0, 0.0, 10.0))
526
- except (StopIteration, KeyError, TypeError):
527
- score = None
528
- scores.append(score)
529
- return scores
530
-
531
- async def _process_waifu_scorer(self, batch_images: list[Image.Image]) -> list[float | None]:
532
- model = self.models.get("waifu_scorer")
533
- if not model: return [None] * len(batch_images)
534
- raw_scores = model(batch_images)
535
- return [float(np.clip(s, 0.0, 10.0)) if s is not None else None for s in raw_scores]
536
-
537
- async def _process_aesthetic_predictor_v2_5(self, batch_images: list[Image.Image]) -> list[float | None]:
538
- model = self.models.get("aesthetic_predictor_v2_5")
539
- if not model: return [None] * len(batch_images)
540
- # Already returns clipped & rounded scores or Nones
541
- return model.inference(batch_images)
542
-
543
- async def _process_anime_aesthetic(self, batch_images: list[Image.Image]) -> list[float | None]:
544
- model = self.models.get("anime_aesthetic")
545
- if not model: return [None] * len(batch_images)
546
-
547
- input_data = preprocess_anime_aesthetic_batch(batch_images)
548
- if input_data is None:
549
- return [None] * len(batch_images)
550
-
551
  try:
552
- preds = model.run(None, {"img": input_data})[0] # Assuming output is (N, 1) or (N,)
553
- scores = [float(np.clip(p.item() * 10.0, 0.0, 10.0)) for p in preds]
554
- return scores
 
 
 
 
 
 
 
 
555
  except Exception as e:
556
- logger.error(f"Error during Anime Aesthetic ONNX prediction: {e}")
557
- return [None] * len(batch_images)
 
558
 
559
- def add_temp_file_for_cleanup(self, file_path: str):
560
- self._temp_files_to_clean.append(file_path)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
561
 
562
- async def shutdown_worker(self):
563
- if self.worker_task and not self.worker_task.done():
564
- logger.info("Attempting to shutdown worker...")
565
- await self.processing_queue.put(None) # Send shutdown signal
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
566
  try:
567
- await asyncio.wait_for(self.worker_task, timeout=5.0)
568
- logger.info("Worker task finished.")
569
- except asyncio.TimeoutError:
570
- logger.warning("Worker task did not finish in time. Cancelling...")
571
- self.worker_task.cancel()
572
  except Exception as e:
573
- logger.error(f"Exception during worker shutdown: {e}")
574
- await self.processing_queue.join() # Wait for queue to be fully processed
575
- logger.info("Processing queue joined.")
576
- self.worker_task = None
577
-
578
-
579
- def cleanup(self):
580
- logger.info("Running cleanup...")
581
- # Shut down asyncio worker
582
- if self.worker_task:
583
- # If running in a context where an event loop is already running
584
- if asyncio.get_event_loop().is_running():
585
- asyncio.create_task(self.shutdown_worker()) # schedule it
586
- else: # If no loop, run it
587
- try:
588
- asyncio.run(self.shutdown_worker())
589
- except RuntimeError as e: # Handles "cannot be called when another loop is running"
590
- logger.error(f"RuntimeError during cleanup's shutdown_worker: {e}. May need manual loop management.")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
591
 
592
- # Clean up temporary CSV files
593
- for f_path in self_temp_files_to_clean:
594
- try:
595
- os.remove(f_path)
596
- logger.info(f"Removed temp file: {f_path}")
597
- except OSError as e:
598
- logger.error(f"Error removing temp file {f_path}: {e}")
599
- self._temp_files_to_clean.clear()
600
- logger.info("Cleanup finished.")
601
-
602
-
603
- #####################################
604
- # Interface #
605
- #####################################
606
-
607
- # Initialize ModelManager once
608
- model_manager = ModelManager(cache_dir=".model_cache")
609
-
610
- def create_interface():
611
- # Define model choices based on ModelManager's loaded models
612
- # Filter out models that failed to load
613
- AVAILABLE_MODEL_KEYS = [k for k in model_manager.model_configs.keys() if k in model_manager.models]
614
- AVAILABLE_MODEL_NAMES_MAP = {k: model_manager.model_configs[k]['name'] for k in AVAILABLE_MODEL_KEYS}
615
 
616
- # [(display_name, value_key), ...] for CheckboxGroup
617
- MODEL_CHOICES_FOR_CHECKBOX = [(AVAILABLE_MODEL_NAMES_MAP[k], k) for k in AVAILABLE_MODEL_KEYS]
618
-
619
-
620
- with gr.Blocks(theme=gr.themes.Soft(primary_hue=gr.themes.colors.blue, secondary_hue=gr.themes.colors.sky)) as demo:
621
- gr.Markdown("""
622
- # Comprehensive Image Evaluation Tool (Refactored)
623
- Upload images to evaluate them using multiple aesthetic and quality prediction models.
624
- Results are displayed in a sortable table with image previews.
625
- """)
626
-
627
- # Stores full processing results (list of dicts)
628
- # Dict keys: 'file_name', 'final_score', and all model_keys with their scores
629
- # This state is the source of truth for regenerating table and CSV
630
- results_state = gr.State([])
631
- # Stores current list of selected model keys (e.g., ['waifu_scorer', 'anime_aesthetic'])
632
- selected_models_state = gr.State(AVAILABLE_MODEL_KEYS)
633
- # Stores current log messages as a list
634
- log_messages_state = gr.State([])
635
 
636
  with gr.Row():
637
- with gr.Column(scale=1): # Inputs
638
- input_images = gr.Files(label="Upload Images", file_count="multiple", type="filepath")
 
 
639
 
640
- if not MODEL_CHOICES_FOR_CHECKBOX:
641
- gr.Markdown("## No models loaded successfully. Please check logs.")
642
- model_checkboxes = None # No models, no checkbox
643
- else:
644
- model_checkboxes = gr.CheckboxGroup(
645
- choices=MODEL_CHOICES_FOR_CHECKBOX,
646
- label="Select Models",
647
- value=AVAILABLE_MODEL_KEYS, # Default to all available selected
648
- info="Choose models for evaluation. Final score is an average of selected model scores."
649
- )
650
-
651
- auto_batch_checkbox = gr.Checkbox(label="Automatic Batch Size Detection", value=True)
652
- batch_size_input = gr.Number(label="Manual Batch Size", value=8, minimum=1, precision=0, interactive=False) # Interactive based on auto_batch_checkbox
653
 
654
- process_btn = gr.Button("Evaluate Images", variant="primary", interactive=bool(MODEL_CHOICES_FOR_CHECKBOX))
655
- clear_btn = gr.Button("Clear Results")
656
- download_csv_btn = gr.Button("Download Results as CSV", variant="secondary")
657
-
658
- with gr.Column(scale=3): # Outputs
659
- progress_tracker = gr.Progress(label="Processing Progress")
660
- log_output = gr.Textbox(label="Logs", lines=10, max_lines=20, interactive=False, autoscroll=True)
 
 
 
 
 
661
 
662
- # Initial headers for DataFrame; will be updated dynamically
663
- initial_df_headers = ['Image', 'File Name'] + [AVAILABLE_MODEL_NAMES_MAP[k] for k in AVAILABLE_MODEL_KEYS] + ['Final Score']
664
  results_dataframe = gr.DataFrame(
665
- headers=initial_df_headers,
666
- datatype=['pil'] + ['str'] * (len(initial_df_headers) -1) , # Image + strings for scores
667
- label="Evaluation Results",
668
- interactive=True, # Enables sorting by clicking headers
669
- row_count=(10, "dynamic"),
670
- col_count=(len(initial_df_headers), "fixed"),
671
- wrap=True,
672
  )
673
- # Hidden file component for download trigger
674
- download_file_provider = gr.File(label="Download Link", visible=False)
675
-
676
- # --- Callback Functions ---
677
- def update_batch_size_interactive(auto_detect_enabled: bool):
678
- return gr.Number.update(interactive=not auto_detect_enabled)
679
-
680
- async def handle_process_images_ui(
681
- files_list: list[gr. rýchle.TempFile] | None, # Gradio File objects
682
- auto_batch_flag: bool,
683
- manual_batch_val: int,
684
- selected_model_keys_from_ui: list[str],
685
- # Gradio will pass the gr.Progress instance automatically by type hinting
686
- # Ensure the name 'progress_tracker_instance' matches an output component if you want to update it by dict key
687
- # Otherwise, use the positional argument `progress`
688
- progress_instance: gr.Progress
689
- ):
690
- if not files_list:
691
- yield {
692
- log_output: "No files uploaded. Please select images first.",
693
- progress_tracker: gr.Progress(0.0, "Idle. No files."),
694
- results_dataframe: gr.DataFrame.update(value=None), # Clear table
695
- results_state: [],
696
- selected_models_state: selected_model_keys_from_ui,
697
- log_messages_state: ["No files uploaded. Please select images first."]
698
- }
699
- return
700
 
701
- # Update selected_models_state right away
702
- yield { selected_models_state: selected_model_keys_from_ui, log_messages_state: [] } # Clear logs state
 
703
 
704
- # Convert Gradio TempFile objects to string paths
705
- actual_file_paths = [f.name for f in files_list]
 
 
 
 
 
706
 
707
- current_log_list = [] # Local log accumulator for this run
708
-
709
- # Call the ModelManager's generator
710
- # The progress_instance is implicitly passed by Gradio to this function
711
- # The ModelManager generator will then use it via its own parameter `progress_tracker_instance`
712
 
713
- # Need to get an async generator from model_manager
714
- evaluation_generator = await model_manager.submit_evaluation_request(
715
- actual_file_paths, auto_batch_flag, manual_batch_val, selected_model_keys_from_ui
716
- )
 
 
 
717
 
718
- dataframe_update_value = None
719
- final_results_for_app_state = []
720
-
721
- async for event in evaluation_generator:
722
- outputs_to_yield = {}
723
- if event["type"] == "log_update":
724
- current_log_list = event["messages"]
725
- outputs_to_yield[log_output] = "\n".join(current_log_list)
726
- elif event["type"] == "progress":
727
- # Update progress bar directly using the passed instance
728
- progress_instance(event["value"], desc=event.get("desc"))
729
- elif event["type"] == "batch_size_update":
730
- outputs_to_yield[batch_size_input] = gr.Number.update(value=event["value"])
731
- elif event["type"] == "partial_results_df_rows":
732
- # data is list of lists for DataFrame rows
733
- # selected_model_keys used to generate current headers
734
- dynamic_headers = ['Image', 'File Name'] + \
735
- [AVAILABLE_MODEL_NAMES_MAP[k] for k in event["selected_model_keys"] if k in AVAILABLE_MODEL_NAMES_MAP] + \
736
- ['Final Score']
737
- dataframe_update_value = pd.DataFrame(event["data"], columns=dynamic_headers) if event["data"] else None
738
- outputs_to_yield[results_dataframe] = gr.DataFrame.update(value=dataframe_update_value, headers=dynamic_headers)
739
-
740
- elif event["type"] == "final_results_state":
741
- final_results_for_app_state = event["data"]
742
-
743
- if outputs_to_yield: # Only yield if there's something to update
744
- yield outputs_to_yield
745
 
746
- # Final updates after generator is exhausted
747
- yield {
748
- results_state: final_results_for_app_state,
749
- log_messages_state: current_log_list, # Save final logs
750
- # DataFrame should be up-to-date from the last partial_results_df_rows
751
- }
 
 
752
 
 
 
 
 
 
 
 
753
 
754
- def handle_clear_results_ui():
755
- # Clear files, logs, table, progress, and internal states
756
  return {
757
- input_images: None,
758
- log_output: "Results cleared.",
759
- results_dataframe: gr.DataFrame.update(value=None, headers=initial_df_headers), # Reset with initial headers
760
- progress_tracker: gr.Progress(0.0, "Idle"),
761
- results_state: [],
762
- # selected_models_state: AVAILABLE_MODEL_KEYS, # Optionally reset model selection
763
- batch_size_input: gr.Number.update(value=8), # Reset batch size
764
- log_messages_state: ["Results cleared."]
765
  }
766
-
767
- # Function to re-render DataFrame and update states when model selection changes
768
- def handle_model_selection_or_state_change_ui(
769
- current_selected_keys: list[str],
770
- current_full_results: list[dict]
771
- ):
772
- if not current_full_results: # No data to process
773
- dynamic_headers = ['Image', 'File Name'] + \
774
- [AVAILABLE_MODEL_NAMES_MAP[k] for k in current_selected_keys if k in AVAILABLE_MODEL_NAMES_MAP] + \
775
- ['Final Score']
776
- return {
777
- results_dataframe: gr.DataFrame.update(value=None, headers=dynamic_headers),
778
- selected_models_state: current_selected_keys,
779
- results_state: current_full_results # pass through if empty
780
- }
781
-
782
- new_df_rows = []
783
- updated_full_results = []
784
 
785
- for res_item_dict in current_full_results:
786
- # Recalculate final score based on new selection
787
- scores_to_avg = []
788
- for mk in current_selected_keys:
789
- if mk in res_item_dict and isinstance(res_item_dict[mk], (float, int)):
790
- scores_to_avg.append(res_item_dict[mk])
791
-
792
- new_final_score = None
793
- if scores_to_avg:
794
- new_final_score_val = float(np.mean(scores_to_avg))
795
- new_final_score = float(np.clip(new_final_score_val, 0.0, 10.0))
796
-
797
- # Update the item in results_state
798
- res_item_dict['final_score'] = new_final_score
799
- updated_full_results.append(res_item_dict.copy()) # Store updated item
800
-
801
- # Prepare row for DataFrame
802
- # Find the corresponding image (this assumes images are not stored in results_state, which they aren't)
803
- # For simplicity, we'll need to re-generate thumbnails if we want them in this update path.
804
- # A robust way: results_state stores paths or minimal data to re-fetch/re-create thumbnails.
805
- # Current implementation of `run_evaluation_generator` directly yields DF rows with PIL images.
806
- # If `handle_model_selection_change_ui` is to re-generate the DF from `results_state`,
807
- # `results_state` items would need to include enough info for `Image.open` and `thumbnail`.
808
- # This is a complex part if we want perfect dynamic DF regeneration with images.
809
- # For now, let's assume `results_state` stores `PIL.Image` thumbnails if this path is critical.
810
- # The `run_evaluation_generator` stores dicts without PIL image objects in `all_results_for_state`.
811
- # This means `handle_model_selection_change_ui` cannot easily reconstruct the 'Image' column.
812
- #
813
- # SIMPLIFICATION: When model selection changes, we only update scores in the existing DataFrame
814
- # if possible, or we re-calculate and re-populate. The current code path re-creates rows.
815
- # To do this properly, `results_state` items should perhaps include original image path or cached thumbnail.
816
- #
817
- # Let's make results_state store {'file_path': ..., 'thumbnail_pil': ..., scores...}
818
- # This needs `run_evaluation_generator` to save file_path and thumbnail_pil to `all_results_for_state`.
819
- # Assume `results_state` items now contain 'thumbnail_pil' and other scores.
820
-
821
- # If 'thumbnail_pil' is not in res_item_dict (because it wasn't saved that way), this will fail.
822
- # This path requires results_state to contain PIL image data for the 'Image' column.
823
- # The current 'run_evaluation_generator' does not save PIL images into `all_results_for_state`.
824
- # It only creates them for immediate DataFrame update.
825
- # This function needs to be re-thought if full DF reconstruction with images is needed here.
826
-
827
- # Let's assume results_state IS NOT used to rebuild the image column.
828
- # The change handler for model_checkboxes will mostly affect the *calculation* of final_score
829
- # and *visibility* of columns if we were dynamically adding/removing them.
830
- # Gradio's DataFrame doesn't easily hide/show columns; we change headers and data.
831
-
832
- # Rebuild row for DF:
833
- df_row = [res_item_dict.get('thumbnail_pil_placeholder', "N/A"), res_item_dict['file_name']]
834
- for mk_cfg in AVAILABLE_MODEL_KEYS: # All possible models to maintain column order
835
- if mk_cfg in current_selected_keys: # If this model is currently selected for display
836
- score = res_item_dict.get(mk_cfg)
837
- df_row.append(f"{score:.4f}" if isinstance(score, (float, int)) else "N/A")
838
- # If not selected, this column won't even be in dynamic_headers.
839
- df_row.append(f"{new_final_score:.4f}" if new_final_score is not None else "N/A")
840
- new_df_rows.append(df_row)
841
-
842
- dynamic_headers = ['Image', 'File Name'] + \
843
- [AVAILABLE_MODEL_NAMES_MAP[k] for k in current_selected_keys if k in AVAILABLE_MODEL_NAMES_MAP] + \
844
- ['Final Score']
845
-
846
- import pandas as pd
847
- df_value = pd.DataFrame(new_df_rows, columns=dynamic_headers) if new_df_rows else None
848
-
849
  return {
850
- results_dataframe: gr.DataFrame.update(value=df_value, headers=dynamic_headers),
851
- selected_models_state: current_selected_keys, # Persist the new selection
852
- results_state: updated_full_results # Persist updated scores
853
  }
854
 
855
-
856
- def handle_download_csv_ui(current_full_results: list[dict], current_selected_keys: list[str]):
857
- if not current_full_results:
858
- # Optionally, send a message to log_output if desired using yield
859
- return gr.File.update(value=None, visible=False)
860
-
861
- # Use StringIO to build CSV in memory
862
- csv_output = StringIO()
863
- # Define fieldnames: Filename, selected model scores, Final Score
864
- fieldnames = ['File Name'] + \
865
- [AVAILABLE_MODEL_NAMES_MAP[k] for k in current_selected_keys if k in AVAILABLE_MODEL_NAMES_MAP] + \
866
- ['Final Score']
867
 
868
- writer = csv.DictWriter(csv_output, fieldnames=fieldnames, extrasaction='ignore')
869
- writer.writeheader()
870
-
871
- for res_item in current_full_results:
872
- row_to_write = {'File Name': res_item['file_name']}
873
- final_score_val = res_item.get('final_score') # This should be up-to-date from results_state
874
- row_to_write['Final Score'] = f"{final_score_val:.4f}" if final_score_val is not None else "N/A"
875
-
876
- for key in current_selected_keys:
877
- if key in AVAILABLE_MODEL_NAMES_MAP: # ensure it's a valid model key
878
- model_display_name = AVAILABLE_MODEL_NAMES_MAP[key]
879
- score_val = res_item.get(key)
880
- row_to_write[model_display_name] = f"{score_val:.4f}" if isinstance(score_val, (float, int)) else "N/A"
881
- writer.writerow(row_to_write)
882
 
883
- csv_content = csv_output.getvalue()
884
- csv_output.close()
885
-
886
- # Save to a temporary file that Gradio can serve
887
- with tempfile.NamedTemporaryFile(mode="w+", delete=False, suffix=".csv", encoding='utf-8') as tmp_file:
888
- tmp_file.write(csv_content)
889
- temp_file_path = tmp_file.name
 
 
 
890
 
891
- model_manager.add_temp_file_for_cleanup(temp_file_path) # Register for cleanup
892
-
893
- return gr.File.update(value=temp_file_path, visible=True, label="results.csv")
894
-
895
-
896
- # --- Wire up components ---
897
- auto_batch_checkbox.change(
898
- fn=update_batch_size_interactive,
899
- inputs=[auto_batch_checkbox],
900
- outputs=[batch_size_input]
901
- )
902
 
903
- # Check if model_checkboxes exists (i.e., models loaded)
904
- if model_checkboxes:
905
- process_btn.click(
906
- fn=handle_process_images_ui,
907
- inputs=[input_images, auto_batch_checkbox, batch_size_input, model_checkboxes],
908
- outputs=[
909
- log_output, progress_tracker, results_dataframe, batch_size_input,
910
- results_state, selected_models_state, log_messages_state # Ensure all yielded components are listed
911
- ]
912
- )
913
- # When model selection changes, update the displayed table and internal states
914
- model_checkboxes.change(
915
- fn=handle_model_selection_or_state_change_ui,
916
- inputs=[model_checkboxes, results_state], # Takes current selection and full results data
917
- outputs=[results_dataframe, selected_models_state, results_state]
918
- )
919
 
920
- clear_btn.click(
921
- fn=handle_clear_results_ui,
922
- outputs=[
923
- input_images, log_output, results_dataframe, progress_tracker,
924
- results_state, batch_size_input, log_messages_state # model_checkboxes could be reset too if needed
925
- ]
926
- )
927
 
928
- download_csv_btn.click(
929
- fn=handle_download_csv_ui,
930
- inputs=[results_state, selected_models_state], # Use current results and selected models for CSV
931
- outputs=[download_file_provider]
932
  )
933
-
934
- # Initial setup on demo load
935
- async def initial_load_setup():
936
- await model_manager.start_worker_if_not_running() # Start async worker
937
- # Set initial state for selected_models_state based on default checkbox values
938
- # This is a bit of a workaround if direct binding isn't available for initial state from component value
939
- return {selected_models_state: AVAILABLE_MODEL_KEYS, log_messages_state: ["Application loaded. Ready."]}
940
-
941
- demo.load(
942
- fn=initial_load_setup,
943
- outputs=[selected_models_state, log_messages_state]
944
  )
945
- # Register cleanup function
946
- demo.unload(model_manager.cleanup)
947
-
948
-
949
- gr.Markdown("""
950
- ### Notes
951
- - **Model Selection**: Dynamically choose models for evaluation. The 'Final Score' and displayed columns update accordingly.
952
- - **Native Table**: Results are shown in a native Gradio DataFrame, allowing sorting by clicking column headers.
953
- - **Batching**: Automatic batch size detection is enabled by default. You can switch to manual batch sizing.
954
- - **CSV Export**: Download the current results (respecting selected models for columns) as a CSV file.
955
- - **Asynchronous Processing**: Image evaluation runs in the background, providing live updates for logs and progress.
956
- """)
957
- return demo
958
 
 
 
 
959
 
960
  if __name__ == "__main__":
961
- # For proper MLP Safetensor loading with adjusted keys, ensure 'safetensors' is installed.
962
- # For WaifuScorer, ensure 'openai-clip' is installed.
963
- # For ONNX models, 'onnxruntime' or 'onnxruntime-gpu'.
964
- # For general Hugging Face model loading, 'transformers'.
965
- # OpenCV ('cv2') for image manipulation: 'opencv-python'.
966
- # And of course 'torch', 'numpy', 'Pillow', 'gradio'.
967
-
968
- # Create a dummy aesthetic_predictor_v2_5.py if it doesn't exist for the stub to work
969
- # (or ensure the real one is present)
970
- if not Path("aesthetic_predictor_v2_5.py").exists():
971
- stub_content = """
972
- # Placeholder for aesthetic_predictor_v2_5.py
973
- # This file needs to contain the actual 'convert_v2_5_from_siglip' function.
974
- # The main script uses a basic stub if this file is missing or fails to import.
975
- # print("aesthetic_predictor_v2_5.py placeholder executed")
976
- def convert_v2_5_from_siglip(*args, **kwargs):
977
- raise NotImplementedError("This is a placeholder. Implement convert_v2_5_from_siglip here or ensure the main script's stub is used.")
978
- """
979
- # Only write if you are sure, or better, let user handle this dependency.
980
- # For this exercise, we assume the main script's internal stub is sufficient if the file is missing.
981
- pass
982
-
983
-
984
- # It's important that the ModelManager is initialized before create_interface() is called,
985
- # as create_interface() relies on model_manager.model_configs.
986
- # This is already handled by placing `model_manager = ModelManager()` globally.
987
-
988
- app_interface = create_interface()
989
- app_interface.queue().launch(debug=True, share=False) # Enable queue for async operations
990
-
991
- # Ensure cleanup is called on exit if demo.unload isn't fully effective in all environments
992
- import atexit
993
- atexit.register(model_manager.cleanup)
 
1
  import os
2
+ import io
3
  import tempfile
4
+ import shutil # Kept for potential future use, but not actively used for now.
 
 
 
 
5
 
6
  import cv2
7
  import numpy as np
8
+ import pandas as pd
9
  import torch
10
  import onnxruntime as rt
11
  from PIL import Image
12
  import gradio as gr
13
+ from transformers import pipeline
14
  from huggingface_hub import hf_hub_download
15
 
16
+ # Assuming aesthetic_predictor_v2_5.py is in the same directory or Python path.
17
+ # If it's not available, the AestheticPredictorV25 model will fail to load.
18
+ # For this example, a mock will be used if the real import fails.
19
+ try:
20
+ from aesthetic_predictor_v2_5 import convert_v2_5_from_siglip
21
+ except ImportError:
22
+ print("Warning: aesthetic_predictor_v2_5.py not found. Using a mock for AestheticPredictorV25.")
23
+ def convert_v2_5_from_siglip(low_cpu_mem_usage=True, trust_remote_code=True):
24
+ # This is a mock.
25
+ mock_model_output = torch.randn(1, 1) # Represents logits for a single image
26
+
27
+ class MockModel(torch.nn.Module):
28
+ def __init__(self):
29
+ super().__init__()
30
+ self.dummy_param = torch.nn.Parameter(torch.empty(0)) # To have a device property
31
+
32
+ def forward(self, pixel_values):
33
+ # Return something that has .logits
34
+ # Batch size from pixel_values
35
+ batch_size = pixel_values.size(0)
36
+ # Create a namedtuple or simple class to mimic HuggingFace output object with .logits
 
 
 
 
 
 
 
 
 
37
  class Output:
38
+ pass
39
+ output = Output()
40
+ output.logits = torch.randn(batch_size, 1).to(self.dummy_param.device)
41
+ return output
42
+
43
+ def to(self, device_or_dtype): # Simplified .to()
44
+ if isinstance(device_or_dtype, torch.dtype):
45
+ # In a real scenario, handle dtype conversion
46
+ pass
47
+ elif isinstance(device_or_dtype, str) or isinstance(device_or_dtype, torch.device):
48
+ self.dummy_param = torch.nn.Parameter(torch.empty(0, device=device_or_dtype)) # Move dummy param to device
49
+ return self
50
+
51
+ def cuda(self): # Mock .cuda()
52
+ return self.to(torch.device('cuda'))
53
+
54
+
55
+ mock_model_instance = MockModel()
56
+
57
+ # Mock preprocessor that returns a dict with "pixel_values"
58
+ mock_preprocessor = lambda images, return_tensors: {"pixel_values": torch.randn(len(images) if isinstance(images, list) else 1, 3, 224, 224)}
59
+ return mock_model_instance, mock_preprocessor
60
+
61
+ # --- Configuration ---
62
+ DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
63
+ DTYPE_WAIFU = torch.float32 # Specific dtype for WaifuScorer's MLP
64
+ CACHE_DIR = None # Set to a path string to use a specific Hugging Face cache directory, e.g., "./hf_cache"
65
+
66
+ # --- Model Definitions ---
67
 
68
  class MLP(torch.nn.Module):
69
+ """Custom MLP for WaifuScorer."""
70
  def __init__(self, input_size: int, batch_norm: bool = True):
71
  super().__init__()
72
  self.input_size = input_size
73
+ self.layers = torch.nn.Sequential(
74
  torch.nn.Linear(self.input_size, 2048), torch.nn.ReLU(),
75
  torch.nn.BatchNorm1d(2048) if batch_norm else torch.nn.Identity(), torch.nn.Dropout(0.3),
76
  torch.nn.Linear(2048, 512), torch.nn.ReLU(),
 
81
  torch.nn.BatchNorm1d(128) if batch_norm else torch.nn.Identity(), torch.nn.Dropout(0.1),
82
  torch.nn.Linear(128, 32), torch.nn.ReLU(),
83
  torch.nn.Linear(32, 1)
84
+ )
85
+ def forward(self, x: torch.Tensor) -> torch.Tensor: return self.layers(x)
 
 
 
 
86
 
87
+ class BaseImageScorer:
88
+ """Abstract base class for image scorers."""
89
+ def __init__(self, model_key: str, model_display_name: str, device: str = DEVICE, verbose: bool = False):
90
+ self.model_key = model_key
91
+ self.model_display_name = model_display_name
92
  self.device = device
93
+ self.verbose = verbose
94
+ self.model = None
95
+ self.preprocessor = None
96
+ self._load_model()
 
 
 
 
 
 
 
97
 
98
+ def _load_model(self): raise NotImplementedError
99
+ def predict(self, images: list[Image.Image]) -> list[float | None]: raise NotImplementedError
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100
 
101
+ def __call__(self, images: list[Image.Image]) -> list[float | None]:
102
+ if not self.model:
103
+ if self.verbose: print(f"{self.model_display_name} model not loaded.")
104
+ return [None] * len(images)
105
+
106
+ rgb_images = [img.convert("RGB") if img.mode != "RGB" else img for img in images]
107
+ return self.predict(rgb_images)
108
 
109
+ class WaifuScorerModel(BaseImageScorer):
110
+ def _load_model(self):
111
+ try:
112
+ import clip
113
+ model_hf_path = "Eugeoter/waifu-scorer-v3/model.pth" # Default path
114
+
115
+ repo_id, filename = os.path.split(model_hf_path)
116
+ actual_model_path = hf_hub_download(repo_id=repo_id, filename=filename, cache_dir=CACHE_DIR)
117
+ if self.verbose: print(f"Loading WaifuScorer MLP from: {actual_model_path}")
118
 
119
+ self.mlp = MLP(input_size=768) # ViT-L/14 embedding size
120
+ if actual_model_path.endswith(".safetensors"):
121
  from safetensors.torch import load_file
122
+ state_dict = load_file(actual_model_path, device=self.device)
123
  else:
124
+ state_dict = torch.load(actual_model_path, map_location=self.device)
 
 
 
 
 
125
  self.mlp.load_state_dict(state_dict)
126
+ self.mlp.to(self.device).eval()
 
127
 
128
+ if self.verbose: print("Loading CLIP model ViT-L/14 for WaifuScorer.")
129
+ self.model, self.preprocessor = clip.load("ViT-L/14", device=self.device) # self.model is CLIP model
130
+ self.model.eval()
131
  except ImportError:
132
+ if self.verbose: print("CLIP library not found. WaifuScorer will not be available.")
133
  except Exception as e:
134
+ if self.verbose: print(f"Error loading WaifuScorer ({self.model_display_name}): {e}")
135
 
136
  @torch.no_grad()
137
+ def predict(self, images: list[Image.Image]) -> list[float | None]:
138
+ if not self.model or not self.mlp: return [None] * len(images)
 
139
 
 
 
 
140
  original_n = len(images)
141
+ processed_images = list(images)
142
+ if original_n == 1: processed_images.append(images[0]) # Duplicate for single image batch
143
 
144
  try:
145
+ image_tensors = torch.cat([self.preprocessor(img).unsqueeze(0) for img in processed_images]).to(self.device)
146
+ image_features = self.model.encode_image(image_tensors)
 
 
147
  norm = image_features.norm(p=2, dim=-1, keepdim=True)
148
+ norm[norm == 0] = 1e-6 # Avoid division by zero, use small epsilon
149
+ im_emb = (image_features / norm).to(device=self.device, dtype=DTYPE_WAIFU)
150
 
151
  predictions = self.mlp(im_emb)
152
+ scores = predictions.clamp(0, 10).cpu().numpy().flatten().tolist()
153
  return scores[:original_n]
154
  except Exception as e:
155
+ if self.verbose: print(f"Error during {self.model_display_name} prediction: {e}")
156
  return [None] * original_n
157
 
158
+ class AestheticPredictorV25(BaseImageScorer):
159
+ def _load_model(self):
160
+ try:
161
+ if self.verbose: print(f"Loading {self.model_display_name}...")
162
+ self.model, self.preprocessor = convert_v2_5_from_siglip(low_cpu_mem_usage=True, trust_remote_code=True)
163
+ # Model's .to() method should handle dtype (e.g. bfloat16) and device.
164
+ self.model = self.model.to(self.device)
165
+ if self.device == 'cuda' and torch.cuda.is_available() and hasattr(self.model, 'to'): # some models might need explicit dtype
166
+ self.model = self.model.to(torch.bfloat16)
167
+ self.model.eval()
168
+ except Exception as e:
169
+ if self.verbose: print(f"Error loading {self.model_display_name}: {e}")
170
 
171
  @torch.no_grad()
172
+ def predict(self, images: list[Image.Image]) -> list[float | None]:
173
+ if not self.model or not self.preprocessor: return [None] * len(images)
 
174
  try:
175
+ inputs = self.preprocessor(images=images, return_tensors="pt")
176
+ pixel_values = inputs["pixel_values"].to(self.model.dummy_param.device if hasattr(self.model, 'dummy_param') else self.device) # Use model's device
177
+ if self.device == 'cuda' and torch.cuda.is_available() and pixel_values.dtype != torch.bfloat16 : # Match dtype if model changed it
178
+ pixel_values = pixel_values.to(torch.bfloat16)
179
+
180
+ output = self.model(pixel_values)
181
+ scores_tensor = output.logits if hasattr(output, 'logits') else output
182
+ scores = scores_tensor.squeeze().float().cpu().numpy()
183
 
184
+ scores_list = [float(np.round(np.clip(s, 0.0, 10.0), 4)) for s in np.atleast_1d(scores)]
185
+ return scores_list
 
 
 
 
186
  except Exception as e:
187
+ if self.verbose: print(f"Error during {self.model_display_name} prediction: {e}")
188
  return [None] * len(images)
189
 
190
+ class AnimeAestheticONNX(BaseImageScorer):
191
+ def _load_model(self):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
192
  try:
193
+ if self.verbose: print(f"Loading {self.model_display_name} (ONNX)...")
194
+ model_path = hf_hub_download(repo_id="skytnt/anime-aesthetic", filename="model.onnx", cache_dir=CACHE_DIR)
195
+ providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if self.device == 'cuda' else ['CPUExecutionProvider']
196
+ valid_providers = [p for p in providers if p in rt.get_available_providers()] or ['CPUExecutionProvider']
197
+ self.model = rt.InferenceSession(model_path, providers=valid_providers)
198
+ if self.verbose: print(f"{self.model_display_name} loaded with providers: {self.model.get_providers()}")
 
199
  except Exception as e:
200
+ if self.verbose: print(f"Error loading {self.model_display_name}: {e}")
201
+
202
+ def _preprocess_image(self, img: Image.Image) -> np.ndarray:
203
+ img_np = np.array(img).astype(np.float32) / 255.0
204
+ s = 768
205
+ h, w = img_np.shape[:2]
206
+ r = min(s/h, s/w)
207
+ new_h, new_w = int(h*r), int(w*r)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
208
 
209
+ resized = cv2.resize(img_np, (new_w, new_h), interpolation=cv2.INTER_AREA if r < 1 else cv2.INTER_LANCZOS4)
 
 
 
 
210
 
211
+ canvas = np.zeros((s, s, 3), dtype=np.float32) # Fill with black
212
+ pad_h, pad_w = (s - new_h) // 2, (s - new_w) // 2
213
+ canvas[pad_h:pad_h+new_h, pad_w:pad_w+new_w] = resized
214
+ return np.transpose(canvas, (2, 0, 1))[np.newaxis, :]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
215
 
216
+ def predict(self, images: list[Image.Image]) -> list[float | None]:
217
+ if not self.model: return [None] * len(images)
218
+ scores = []
219
+ for img in images:
 
 
 
 
 
 
 
 
 
 
 
220
  try:
221
+ input_tensor = self._preprocess_image(img)
222
+ pred = self.model.run(None, {"img": input_tensor})[0].item()
223
+ scores.append(float(np.clip(pred * 10.0, 0.0, 10.0)))
 
 
224
  except Exception as e:
225
+ if self.verbose: print(f"Error predicting with {self.model_display_name} for one image: {e}")
226
+ scores.append(None)
227
+ return scores
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
228
 
229
+ class AestheticShadowPipeline(BaseImageScorer):
230
+ def _load_model(self):
231
+ try:
232
+ if self.verbose: print(f"Loading {self.model_display_name} pipeline...")
233
+ pipeline_device = 0 if self.device == 'cuda' else -1
234
+ self.model = pipeline("image-classification", model="NeoChen1024/aesthetic-shadow-v2-backup", device=pipeline_device)
235
+ except Exception as e:
236
+ if self.verbose: print(f"Error loading {self.model_display_name}: {e}")
237
 
238
+ def predict(self, images: list[Image.Image]) -> list[float | None]:
239
+ if not self.model: return [None] * len(images)
 
 
240
  scores = []
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
241
  try:
242
+ pipeline_results = self.model(images, top_k=None) # Assuming pipeline handles batching
243
+
244
+ # Ensure consistent output structure from pipeline (List[List[Dict]] vs List[Dict])
245
+ if images and pipeline_results and not isinstance(pipeline_results[0], list):
246
+ pipeline_results = [pipeline_results]
247
+
248
+ for res_set in pipeline_results:
249
+ try:
250
+ hq_score_dict = next(p for p in res_set if p['label'] == 'hq')
251
+ scores.append(float(np.clip(hq_score_dict['score'] * 10.0, 0.0, 10.0)))
252
+ except (StopIteration, TypeError, KeyError): scores.append(None)
253
  except Exception as e:
254
+ if self.verbose: print(f"Error during {self.model_display_name} prediction: {e}")
255
+ return [None] * len(images) # All None if batch fails
256
+ return scores
257
 
258
+ # --- Model Management ---
259
+ MODEL_REGISTRY = {
260
+ "aesthetic_shadow": {"class": AestheticShadowPipeline, "name": "Aesthetic Shadow"},
261
+ "waifu_scorer": {"class": WaifuScorerModel, "name": "Waifu Scorer"},
262
+ "aesthetic_predictor_v2_5": {"class": AestheticPredictorV25, "name": "Aesthetic V2.5"},
263
+ "anime_aesthetic": {"class": AnimeAestheticONNX, "name": "Anime Score"},
264
+ }
265
+ LOADED_MODELS = {} # Populated at startup
266
+
267
+ def initialize_models(verbose_loading=False):
268
+ print(f"Using device: {DEVICE}")
269
+ print("Initializing models...")
270
+ for key, config in MODEL_REGISTRY.items():
271
+ LOADED_MODELS[key] = config["class"](key, config['name'], device=DEVICE, verbose=verbose_loading)
272
+ print("Model initialization complete.")
273
+
274
+ # --- Core Logic ---
275
+ @torch.no_grad()
276
+ def auto_tune_batch_size(images: list[Image.Image], selected_model_keys: list[str],
277
+ initial_bs: int = 1, max_bs_limit: int = 64, verbose: bool = False) -> int:
278
+ if not images or not selected_model_keys: return initial_bs
279
+ if verbose: print("Auto-tuning batch size...")
280
+
281
+ test_image = images[0]
282
+ active_models = [LOADED_MODELS[key] for key in selected_model_keys if key in LOADED_MODELS and LOADED_MODELS[key].model]
283
+ if not active_models: return initial_bs
284
 
285
+ bs = initial_bs
286
+ optimal_bs = initial_bs
287
+ while bs <= len(images) and bs <= max_bs_limit:
288
+ try:
289
+ batch_test_images = [test_image] * bs
290
+ for model in active_models:
291
+ if verbose: print(f" Testing {model.model_display_name} with batch size {bs}")
292
+ model.predict(batch_test_images)
293
+ if DEVICE == 'cuda': torch.cuda.empty_cache()
294
+
295
+ optimal_bs = bs
296
+ if bs == max_bs_limit: break
297
+ bs = min(bs * 2, max_bs_limit) # Try next power of 2 or max_bs_limit
298
+ except Exception as e: # Typically OOM or other runtime errors
299
+ if verbose: print(f" Failed at batch size {bs} ({type(e).__name__}). Optimal so far: {optimal_bs}. Error: {str(e)[:100]}")
300
+ break
301
+ if verbose: print(f"Auto-tuned batch size: {optimal_bs}")
302
+ return max(1, optimal_bs)
303
+
304
+ async def evaluate_images_core(
305
+ pil_images: list[Image.Image], file_names: list[str],
306
+ selected_model_keys: list[str], batch_size: int,
307
+ progress_tracker: gr.Progress
308
+ ) -> tuple[pd.DataFrame, list[str]]:
309
+
310
+ logs = []
311
+ num_images = len(pil_images)
312
+ if num_images == 0: return pd.DataFrame(), ["No images to process."]
313
+
314
+ # Initialize results_data: list of dicts, one per image
315
+ results_data = [{'File Name': fn, 'Thumbnail': img.copy().resize((150,150)), 'Final Score': np.nan}
316
+ for fn, img in zip(file_names, pil_images)]
317
+ for r_dict in results_data: # Initialize all model score columns to NaN
318
+ for cfg in MODEL_REGISTRY.values(): r_dict[cfg['name']] = np.nan
319
+
320
+ progress_tracker(0, desc="Starting evaluation...")
321
+ total_models_to_run = len(selected_model_keys)
322
+
323
+ for model_idx, model_key in enumerate(selected_model_keys):
324
+ model = LOADED_MODELS.get(model_key)
325
+ if not model or not model.model:
326
+ logs.append(f"Skipping {MODEL_REGISTRY[model_key]['name']} (not loaded).")
327
+ continue
328
+
329
+ model_name = model.model_display_name
330
+ logs.append(f"Processing with {model_name}...")
331
+
332
+ current_img_offset = 0
333
+ for batch_start_idx in range(0, num_images, batch_size):
334
+ # Progress: (current_model_idx + fraction_of_current_model_done) / total_models_to_run
335
+ model_progress_fraction = (batch_start_idx / num_images)
336
+ overall_progress = (model_idx + model_progress_fraction) / total_models_to_run
337
+ progress_tracker(overall_progress, desc=f"{model_name} (Batch {batch_start_idx//batch_size + 1})")
338
+
339
+ batch_images = pil_images[batch_start_idx : batch_start_idx + batch_size]
340
  try:
341
+ scores = model(batch_images) # Use __call__
342
+ for i, score in enumerate(scores):
343
+ results_data[current_img_offset + i][model_name] = score if score is not None else np.nan
 
 
344
  except Exception as e:
345
+ logs.append(f"Error with {model_name} on batch: {e}")
346
+ current_img_offset += len(batch_images)
347
+ logs.append(f"Finished with {model_name}.")
348
+
349
+ # Calculate Final Scores
350
+ for i in range(num_images):
351
+ img_scores = [results_data[i][MODEL_REGISTRY[mk]['name']] for mk in selected_model_keys
352
+ if pd.notna(results_data[i].get(MODEL_REGISTRY[mk]['name']))]
353
+ if img_scores:
354
+ results_data[i]['Final Score'] = float(np.clip(np.mean(img_scores), 0.0, 10.0))
355
+
356
+ df = pd.DataFrame(results_data)
357
+ # Define column order: Thumbnail, File Name, then model scores, then Final Score
358
+ ordered_cols = ['Thumbnail', 'File Name'] + \
359
+ [MODEL_REGISTRY[k]['name'] for k in MODEL_REGISTRY.keys() if MODEL_REGISTRY[k]['name'] in df.columns] + \
360
+ ['Final Score']
361
+ df = df[[col for col in ordered_cols if col in df.columns]] # Ensure all columns exist
362
+
363
+ logs.append("Evaluation complete.")
364
+ progress_tracker(1.0, desc="Evaluation complete.")
365
+ return df, logs
366
+
367
+ def results_df_to_csv_bytes(df: pd.DataFrame, selected_model_display_names: list[str]) -> bytes | None:
368
+ if df.empty: return None
369
+
370
+ cols_for_csv = ['File Name', 'Final Score'] + \
371
+ [name for name in selected_model_display_names if name in df.columns and name not in cols_for_csv]
372
+
373
+ df_csv = df[cols_for_csv].copy()
374
+ for col in df_csv.select_dtypes(include=['float']).columns: # Format float scores
375
+ df_csv[col] = df_csv[col].apply(lambda x: f"{x:.4f}" if pd.notnull(x) else "N/A")
376
+
377
+ s_io = io.StringIO()
378
+ df_csv.to_csv(s_io, index=False)
379
+ return s_io.getvalue().encode('utf-8')
380
 
381
+ # --- Gradio Interface ---
382
+ def create_gradio_interface():
383
+ model_name_choices = [config['name'] for config in MODEL_REGISTRY.values()]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
384
 
385
+ # Define column structure for DataFrame
386
+ initial_df_cols = ['Thumbnail', 'File Name'] + model_name_choices + ['Final Score']
387
+ initial_datatypes = ['image', 'str'] + ['number'] * (len(model_name_choices) + 1)
388
+
389
+ with gr.Blocks(theme=gr.themes.Glass()) as demo:
390
+ gr.Markdown("## ✨ Comprehensive Image Evaluation Tool ✨")
391
+
392
+ # For storing results DataFrame between interactions
393
+ results_state = gr.State(pd.DataFrame(columns=initial_df_cols))
 
 
 
 
 
 
 
 
 
 
394
 
395
  with gr.Row():
396
+ with gr.Column(scale=1, min_width=300):
397
+ gr.Markdown("#### Controls")
398
+ files_input = gr.Files(label="Upload Images", file_count="multiple", type="filepath")
399
+ models_checkbox_group = gr.CheckboxGroup(choices=model_name_choices, value=model_name_choices, label="Select Models")
400
 
401
+ with gr.Accordion("Batch Settings", open=False):
402
+ auto_batch_toggle = gr.Checkbox(label="Auto-detect Batch Size", value=True)
403
+ manual_batch_input = gr.Number(label="Manual Batch Size", value=4, minimum=1, step=1, interactive=False) # Interactive based on toggle
 
 
 
 
 
 
 
 
 
 
404
 
405
+ evaluate_button = gr.Button("🚀 Evaluate Images", variant="primary")
406
+ with gr.Row():
407
+ clear_button = gr.Button("🧹 Clear")
408
+ download_button = gr.Button("💾 Download CSV")
409
+
410
+ # Hidden component for file download functionality
411
+ csv_file_output = gr.File(label="Download CSV File", visible=False)
412
+
413
+ with gr.Column(scale=3, min_width=600):
414
+ gr.Markdown("#### Results")
415
+ # Using gr.Slider for progress display
416
+ progress_slider = gr.Slider(label="Progress", minimum=0, maximum=1, value=0, interactive=False)
417
 
 
 
418
  results_dataframe = gr.DataFrame(
419
+ label="Evaluation Scores",
420
+ headers=initial_df_cols,
421
+ datatype=initial_datatypes,
422
+ interactive=True, # Enables native sorting by clicking headers
423
+ height=500,
424
+ wrap=True
 
425
  )
426
+ logs_textbox = gr.Textbox(label="Process Logs", lines=5, max_lines=10, interactive=False)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
427
 
428
+ # --- Callbacks ---
429
+ def map_display_names_to_keys(display_names: list[str]) -> list[str]:
430
+ return [key for key, cfg in MODEL_REGISTRY.items() if cfg['name'] in display_names]
431
 
432
+ async def run_evaluation(uploaded_files, selected_model_names, auto_batch, manual_batch,
433
+ current_results_df, progress=gr.Progress(track_tqdm=True)):
434
+ if not uploaded_files:
435
+ return {
436
+ results_state: current_results_df, logs_textbox: "No files uploaded. Please upload images first.",
437
+ progress_slider: gr.update(value=0, label="Progress")
438
+ }
439
 
440
+ yield {logs_textbox: "Loading images...", progress_slider: gr.update(value=0.01, label="Loading images...")}
 
 
 
 
441
 
442
+ pil_images, file_names = [], []
443
+ for f_obj in uploaded_files:
444
+ try:
445
+ pil_images.append(Image.open(f_obj.name).convert("RGB")) # f_obj.name is path for type="filepath"
446
+ file_names.append(os.path.basename(f_obj.name))
447
+ except Exception as e:
448
+ print(f"Error loading image {f_obj.name}: {e}") # Log to console
449
 
450
+ if not pil_images:
451
+ return {logs_textbox: "No valid images could be loaded.", progress_slider: gr.update(value=0, label="Error")}
452
+
453
+ selected_keys = map_display_names_to_keys(selected_model_names)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
454
 
455
+ batch_size_to_use = manual_batch
456
+ if auto_batch:
457
+ yield {logs_textbox: "Auto-tuning batch size...", progress_slider: gr.update(value=0.1, label="Auto-tuning...")}
458
+ batch_size_to_use = auto_tune_batch_size(pil_images, selected_keys, verbose=True)
459
+ yield {manual_batch_input: gr.update(value=batch_size_to_use)} # Update UI with detected size
460
+
461
+ yield {logs_textbox: f"Starting evaluation with batch size {batch_size_to_use}...",
462
+ progress_slider: gr.update(value=0.15, label=f"Evaluating (Batch: {batch_size_to_use})...")}
463
 
464
+ df_new_results, log_messages = await evaluate_images_core(
465
+ pil_images, file_names, selected_keys, batch_size_to_use, progress
466
+ )
467
+
468
+ # Sort by 'Final Score' descending by default before display
469
+ if not df_new_results.empty and 'Final Score' in df_new_results.columns:
470
+ df_new_results = df_new_results.sort_values(by='Final Score', ascending=False, na_position='last')
471
 
 
 
472
  return {
473
+ results_state: df_new_results, results_dataframe: df_new_results,
474
+ logs_textbox: "\n".join(log_messages),
475
+ progress_slider: gr.update(value=1.0, label="Evaluation Complete")
 
 
 
 
 
476
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
477
 
478
+ def clear_all_outputs():
479
+ empty_df = pd.DataFrame(columns=initial_df_cols)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
480
  return {
481
+ results_state: empty_df, results_dataframe: empty_df,
482
+ files_input: None, logs_textbox: "Outputs cleared.",
483
+ progress_slider: gr.update(value=0, label="Progress")
484
  }
485
 
486
+ def download_csv_file(current_df, selected_names):
487
+ if current_df.empty:
488
+ gr.Warning("No results available to download.")
489
+ return None
 
 
 
 
 
 
 
 
490
 
491
+ csv_data = results_df_to_csv_bytes(current_df, selected_names)
492
+ if csv_data:
493
+ with tempfile.NamedTemporaryFile(delete=False, suffix=".csv", mode='wb') as tmp_f:
494
+ tmp_f.write(csv_data)
495
+ gr.Info("CSV file prepared for download.")
496
+ return tmp_f.name
497
+ gr.Error("Failed to generate CSV.")
498
+ return None
499
+
500
+ def update_final_scores_on_model_select(selected_model_names, current_df):
501
+ if current_df.empty: return current_df
 
 
 
502
 
503
+ df_updated = current_df.copy()
504
+ selected_keys = map_display_names_to_keys(selected_model_names)
505
+
506
+ for i, row in df_updated.iterrows():
507
+ img_scores = [row[MODEL_REGISTRY[mk]['name']] for mk in selected_keys
508
+ if pd.notna(row.get(MODEL_REGISTRY[mk]['name']))]
509
+ if img_scores:
510
+ df_updated.loc[i, 'Final Score'] = float(np.clip(np.mean(img_scores), 0.0, 10.0))
511
+ else:
512
+ df_updated.loc[i, 'Final Score'] = np.nan
513
 
514
+ if 'Final Score' in df_updated.columns: # Re-sort
515
+ df_updated = df_updated.sort_values(by='Final Score', ascending=False, na_position='last')
 
 
 
 
 
 
 
 
 
516
 
517
+ return {results_state: df_updated, results_dataframe: df_updated}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
518
 
519
+ auto_batch_toggle.change(lambda x: gr.update(interactive=not x), inputs=auto_batch_toggle, outputs=manual_batch_input)
 
 
 
 
 
 
520
 
521
+ evaluate_button.click(
522
+ fn=run_evaluation,
523
+ inputs=[files_input, models_checkbox_group, auto_batch_toggle, manual_batch_input, results_state],
524
+ outputs=[results_state, results_dataframe, logs_textbox, manual_batch_input, progress_slider]
525
  )
526
+ clear_button.click(fn=clear_all_outputs, outputs=[results_state, results_dataframe, files_input, logs_textbox, progress_slider])
527
+ download_button.click(fn=download_csv_file, inputs=[results_state, models_checkbox_group], outputs=csv_file_output)
528
+ models_checkbox_group.change(
529
+ fn=update_final_scores_on_model_select,
530
+ inputs=[models_checkbox_group, results_state],
531
+ outputs=[results_state, results_dataframe]
 
 
 
 
 
532
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
533
 
534
+ # Initial load state for the DataFrame UI component
535
+ demo.load(lambda: pd.DataFrame(columns=initial_df_cols), outputs=[results_dataframe])
536
+ return demo
537
 
538
  if __name__ == "__main__":
539
+ initialize_models(verbose_loading=True) # Load models once at startup
540
+ gradio_app = create_gradio_interface()
541
+ gradio_app.queue().launch(debug=False) # Enable queue for async ops, debug=True for more logs