Update app.py
Browse files
app.py
CHANGED
@@ -12,15 +12,16 @@ import torch
|
|
12 |
import onnxruntime as rt
|
13 |
from PIL import Image
|
14 |
from huggingface_hub import hf_hub_download
|
15 |
-
from transformers import pipeline, Pipeline
|
16 |
from tqdm import tqdm
|
17 |
|
18 |
-
# Suppress a specific PIL warning about image size
|
19 |
Image.MAX_IMAGE_PIXELS = None
|
20 |
|
21 |
# --- Configuration ---
|
22 |
CACHE_DIR = "./hf_cache"
|
23 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
24 |
DTYPE = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8 else torch.float32
|
25 |
|
26 |
print(f"Using device: {DEVICE} with dtype: {DTYPE}")
|
@@ -31,21 +32,20 @@ print(f"Using device: {DEVICE} with dtype: {DTYPE}")
|
|
31 |
|
32 |
class AestheticScorer(ABC):
|
33 |
"""Abstract base class for all aesthetic scoring models."""
|
34 |
-
|
35 |
def __init__(self, model_name: str, repo_id: str, filename: str = None):
|
36 |
self.model_name = model_name
|
37 |
self.repo_id = repo_id
|
38 |
self.filename = filename
|
39 |
self._model = None
|
40 |
-
print(f"Initializing scorer: {self.model_name}")
|
41 |
|
42 |
@property
|
43 |
def model(self):
|
44 |
"""Lazy-loads the model on first access."""
|
45 |
if self._model is None:
|
46 |
-
print(f"Loading model for '{self.model_name}'...")
|
47 |
self._model = self.load_model()
|
48 |
-
print(f"'{self.model_name}' model loaded.")
|
49 |
return self._model
|
50 |
|
51 |
def _download_model(self) -> str:
|
@@ -63,9 +63,9 @@ class AestheticScorer(ABC):
|
|
63 |
pass
|
64 |
|
65 |
def release_model(self):
|
66 |
-
"""Releases model from memory."""
|
67 |
if self._model is not None:
|
68 |
-
print(f"Releasing model: {self.model_name}")
|
69 |
del self._model
|
70 |
self._model = None
|
71 |
gc.collect()
|
@@ -74,23 +74,15 @@ class AestheticScorer(ABC):
|
|
74 |
|
75 |
class PipelineScorer(AestheticScorer):
|
76 |
"""Scorer for models compatible with Hugging Face pipelines."""
|
77 |
-
|
78 |
def load_model(self) -> Pipeline:
|
79 |
-
""
|
80 |
-
return pipeline(
|
81 |
-
"image-classification",
|
82 |
-
model=self.repo_id,
|
83 |
-
device=DEVICE,
|
84 |
-
)
|
85 |
|
86 |
@torch.no_grad()
|
87 |
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
88 |
-
|
89 |
-
results = self.model(image_batch)
|
90 |
scores = []
|
91 |
for res in results:
|
92 |
try:
|
93 |
-
# Find the score for the 'hq' (high quality) label
|
94 |
hq_score = next(item['score'] for item in res if item['label'] == 'hq')
|
95 |
scores.append(round(hq_score * 10.0, 4))
|
96 |
except (StopIteration, TypeError):
|
@@ -99,21 +91,16 @@ class PipelineScorer(AestheticScorer):
|
|
99 |
|
100 |
class ONNXScorer(AestheticScorer):
|
101 |
"""Scorer for ONNX-based models."""
|
102 |
-
|
103 |
def load_model(self) -> rt.InferenceSession:
|
104 |
-
"""Loads an ONNX inference session."""
|
105 |
model_path = self._download_model()
|
106 |
return rt.InferenceSession(model_path, providers=['CUDAExecutionProvider' if DEVICE == 'cuda' else 'CPUExecutionProvider'])
|
107 |
|
108 |
def _preprocess(self, img: Image.Image) -> np.ndarray:
|
109 |
-
"""Preprocesses a single image for the Anime Aesthetic model."""
|
110 |
img_np = np.array(img.convert("RGB")).astype(np.float32) / 255.0
|
111 |
s = 768
|
112 |
h, w = img_np.shape[:2]
|
113 |
-
|
114 |
-
|
115 |
-
else:
|
116 |
-
new_h, new_w = int(s * h / w), s
|
117 |
|
118 |
resized = cv2.resize(img_np, (new_w, new_h), interpolation=cv2.INTER_AREA)
|
119 |
canvas = np.zeros((s, s, 3), dtype=np.float32)
|
@@ -123,7 +110,6 @@ class ONNXScorer(AestheticScorer):
|
|
123 |
return np.transpose(canvas, (2, 0, 1))[np.newaxis, :]
|
124 |
|
125 |
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
126 |
-
"""Scores images one by one as this model doesn't support batching."""
|
127 |
scores = []
|
128 |
for img in image_batch:
|
129 |
try:
|
@@ -135,264 +121,192 @@ class ONNXScorer(AestheticScorer):
|
|
135 |
return scores
|
136 |
|
137 |
class CLIPMLPScorer(AestheticScorer):
|
138 |
-
"""Scorer for models using a CLIP backbone and
|
139 |
-
|
140 |
class MLP(torch.nn.Module):
|
|
|
141 |
def __init__(self, input_size: int):
|
142 |
super().__init__()
|
143 |
self.layers = torch.nn.Sequential(
|
144 |
-
torch.nn.Linear(input_size,
|
145 |
torch.nn.ReLU(),
|
146 |
-
torch.nn.
|
147 |
-
torch.nn.
|
|
|
|
|
|
|
|
|
|
|
148 |
torch.nn.ReLU(),
|
|
|
149 |
torch.nn.Dropout(0.2),
|
150 |
-
torch.nn.Linear(
|
151 |
torch.nn.ReLU(),
|
152 |
-
torch.nn.
|
|
|
|
|
153 |
torch.nn.ReLU(),
|
154 |
-
torch.nn.Linear(
|
155 |
)
|
156 |
def forward(self, x):
|
157 |
return self.layers(x)
|
158 |
|
159 |
def load_model(self) -> Dict[str, Any]:
|
160 |
-
|
161 |
-
import clip # Lazy import
|
162 |
-
|
163 |
model_path = self._download_model()
|
164 |
-
|
165 |
mlp = self.MLP(input_size=768) # ViT-L/14 has 768 features
|
166 |
state_dict = torch.load(model_path, map_location=DEVICE)
|
167 |
mlp.load_state_dict(state_dict)
|
168 |
-
mlp.to(device=DEVICE
|
169 |
mlp.eval()
|
170 |
-
|
171 |
clip_model, preprocess = clip.load("ViT-L/14", device=DEVICE)
|
172 |
-
|
173 |
return {"mlp": mlp, "clip": clip_model, "preprocess": preprocess}
|
174 |
|
175 |
@torch.no_grad()
|
176 |
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
177 |
-
"""Scores a batch using CLIP features and the MLP head."""
|
178 |
preprocess = self.model['preprocess']
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
179 |
image_tensors = torch.cat([preprocess(img).unsqueeze(0) for img in image_batch]).to(DEVICE)
|
180 |
-
|
181 |
-
image_features = self.model['clip'].encode_image(image_tensors)
|
182 |
image_features /= image_features.norm(dim=-1, keepdim=True)
|
|
|
|
|
183 |
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
return [round(float(s), 4) for s in scores]
|
189 |
|
190 |
# --- Model Registry ---
|
191 |
-
MODEL_REGISTRY: Dict[str,
|
192 |
-
"Aesthetic Shadow V2": PipelineScorer(
|
193 |
-
|
194 |
-
),
|
195 |
-
"
|
196 |
-
"Waifu Scorer V2", "skytnt/waifu-aesthetic-scorer", "model.pth"
|
197 |
-
),
|
198 |
-
"Anime Scorer": ONNXScorer(
|
199 |
-
"Anime Scorer", "skytnt/anime-aesthetic", "model.onnx"
|
200 |
-
)
|
201 |
}
|
202 |
-
|
203 |
-
# In-memory cache for loaded model instances
|
204 |
_loaded_models_cache: Dict[str, AestheticScorer] = {}
|
205 |
|
206 |
-
|
207 |
# ==================================================================================
|
208 |
# 2. CORE PROCESSING LOGIC
|
209 |
# ==================================================================================
|
210 |
|
211 |
def get_scorers(model_names: List[str]) -> List[AestheticScorer]:
|
212 |
"""Retrieves and caches scorer instances based on selected names."""
|
213 |
-
|
214 |
-
for name, scorer in list(_loaded_models_cache.items()):
|
215 |
if name not in model_names:
|
216 |
-
|
217 |
del _loaded_models_cache[name]
|
218 |
-
|
219 |
-
# Load newly selected models
|
220 |
-
scorers = []
|
221 |
-
for name in model_names:
|
222 |
-
if name in _loaded_models_cache:
|
223 |
-
scorers.append(_loaded_models_cache[name])
|
224 |
-
elif name in MODEL_REGISTRY:
|
225 |
-
scorer = MODEL_REGISTRY[name]
|
226 |
-
_loaded_models_cache[name] = scorer
|
227 |
-
scorers.append(scorer)
|
228 |
-
return scorers
|
229 |
|
230 |
def evaluate_images(
|
231 |
-
files: List[gr.File],
|
232 |
-
selected_model_names: List[str],
|
233 |
-
batch_size: int,
|
234 |
-
progress: gr.Progress = gr.Progress(track_tqdm=True),
|
235 |
) -> pd.DataFrame:
|
236 |
-
"""
|
237 |
-
Main function to process images, run them through selected models,
|
238 |
-
and return results as a Pandas DataFrame.
|
239 |
-
"""
|
240 |
if not files:
|
241 |
gr.Warning("No images uploaded. Please upload files to evaluate.")
|
242 |
return pd.DataFrame()
|
243 |
-
|
244 |
if not selected_model_names:
|
245 |
gr.Warning("No models selected. Please select at least one model.")
|
246 |
return pd.DataFrame()
|
247 |
|
248 |
try:
|
249 |
image_paths = [Path(f.name) for f in files]
|
250 |
-
all_results = []
|
251 |
-
scorers = get_scorers(selected_model_names)
|
252 |
-
|
253 |
-
# Use a single tqdm instance for progress tracking
|
254 |
-
pbar = tqdm(total=len(image_paths), desc="Processing images")
|
255 |
|
256 |
-
for i in range(0, len(image_paths), batch_size):
|
257 |
batch_paths = image_paths[i : i + batch_size]
|
258 |
-
|
259 |
-
# Load images for the current batch
|
260 |
try:
|
261 |
batch_images = [Image.open(p).convert("RGB") for p in batch_paths]
|
262 |
except Exception as e:
|
263 |
gr.Warning(f"Skipping a batch due to an error loading an image: {e}")
|
264 |
-
pbar.update(len(batch_paths))
|
265 |
continue
|
266 |
-
|
267 |
-
# Get scores from all selected models for the batch
|
268 |
-
batch_scores: Dict[str, List[float]] = {}
|
269 |
-
for scorer in scorers:
|
270 |
-
batch_scores[scorer.model_name] = scorer.score_batch(batch_images)
|
271 |
|
272 |
-
|
|
|
273 |
for j, path in enumerate(batch_paths):
|
274 |
-
result_row = {"Image":
|
275 |
-
|
276 |
-
scores_for_avg = []
|
277 |
for scorer in scorers:
|
278 |
-
|
279 |
-
|
280 |
-
scores_for_avg.append(score)
|
281 |
-
|
282 |
-
# Calculate average score
|
283 |
-
if scores_for_avg:
|
284 |
-
result_row["Average Score"] = round(np.mean(scores_for_avg), 4)
|
285 |
-
else:
|
286 |
-
result_row["Average Score"] = 0.0
|
287 |
-
|
288 |
all_results.append(result_row)
|
289 |
-
|
290 |
-
pbar.update(len(batch_paths))
|
291 |
|
292 |
-
|
293 |
-
|
294 |
-
if not all_results:
|
295 |
-
gr.Warning("Processing completed, but no results were generated.")
|
296 |
-
return pd.DataFrame()
|
297 |
-
|
298 |
-
return pd.DataFrame(all_results)
|
299 |
|
300 |
except Exception as e:
|
301 |
gr.Error(f"A critical error occurred: {e}")
|
302 |
-
# Clean up in case of failure
|
303 |
-
for scorer in _loaded_models_cache.values():
|
304 |
-
scorer.release_model()
|
305 |
-
_loaded_models_cache.clear()
|
306 |
return pd.DataFrame()
|
307 |
|
308 |
-
|
309 |
# ==================================================================================
|
310 |
# 3. GRADIO USER INTERFACE
|
311 |
# ==================================================================================
|
312 |
|
313 |
def create_ui() -> gr.Blocks:
|
314 |
"""Creates and configures the Gradio web interface."""
|
315 |
-
|
316 |
all_model_names = list(MODEL_REGISTRY.keys())
|
317 |
-
|
318 |
-
# Define headers and datatypes for the results table
|
319 |
dataframe_headers = ["Image", "Filename"] + all_model_names + ["Average Score"]
|
320 |
dataframe_datatypes = ["image", "str"] + ["number"] * (len(all_model_names) + 1)
|
321 |
|
322 |
-
with gr.Blocks(theme=gr.themes.Soft(), title="Image Aesthetic Scorer") as demo:
|
323 |
-
gr.Markdown(
|
324 |
-
|
325 |
-
# 🖼️ Modern Image Aesthetic Scorer
|
326 |
-
Upload your images, select the scoring models, and click 'Evaluate'.
|
327 |
-
The results table supports **interactive sorting** (click on headers) and can be **downloaded as a CSV**.
|
328 |
-
"""
|
329 |
-
)
|
330 |
|
331 |
with gr.Row():
|
332 |
with gr.Column(scale=1):
|
333 |
-
gr.
|
334 |
-
|
335 |
-
|
336 |
-
file_count="multiple",
|
337 |
-
file_types=["image"],
|
338 |
-
)
|
339 |
-
|
340 |
-
with gr.Accordion("Advanced Configuration", open=False):
|
341 |
-
model_checkboxes = gr.CheckboxGroup(
|
342 |
-
choices=all_model_names,
|
343 |
-
value=all_model_names,
|
344 |
-
label="Scoring Models",
|
345 |
-
info="Choose which models to use for evaluation.",
|
346 |
-
)
|
347 |
-
batch_size_slider = gr.Slider(
|
348 |
-
minimum=1,
|
349 |
-
maximum=64,
|
350 |
-
value=8,
|
351 |
-
step=1,
|
352 |
-
label="Batch Size",
|
353 |
-
info="Adjust based on your VRAM. Higher is faster.",
|
354 |
-
)
|
355 |
-
|
356 |
with gr.Row():
|
357 |
process_button = gr.Button("🚀 Evaluate Images", variant="primary")
|
358 |
clear_button = gr.Button("🧹 Clear All")
|
359 |
|
360 |
with gr.Column(scale=3):
|
361 |
-
|
362 |
results_dataframe = gr.DataFrame(
|
363 |
headers=dataframe_headers,
|
364 |
datatype=dataframe_datatypes,
|
365 |
label="Evaluation Scores",
|
366 |
interactive=True,
|
367 |
-
|
|
|
368 |
)
|
369 |
-
# This is a cleaner way to show the download button
|
370 |
-
results_dataframe.style(height=800, show_download_button=True)
|
371 |
-
|
372 |
|
373 |
-
# --- Event Handlers ---
|
374 |
process_button.click(
|
375 |
fn=evaluate_images,
|
376 |
inputs=[input_files, model_checkboxes, batch_size_slider],
|
377 |
-
outputs=[results_dataframe]
|
378 |
-
concurrency_limit=1 # Only one evaluation at a time
|
379 |
)
|
380 |
|
381 |
def clear_outputs():
|
382 |
-
|
383 |
-
for scorer in _loaded_models_cache.values():
|
384 |
scorer.release_model()
|
385 |
_loaded_models_cache.clear()
|
386 |
gr.Info("Cleared results and released models from memory.")
|
387 |
-
|
388 |
-
return pd.DataFrame()
|
389 |
-
|
390 |
-
clear_button.click(
|
391 |
-
fn=clear_outputs,
|
392 |
-
inputs=[],
|
393 |
-
outputs=[results_dataframe],
|
394 |
-
)
|
395 |
|
|
|
396 |
return demo
|
397 |
|
398 |
# ==================================================================================
|
@@ -400,8 +314,6 @@ def create_ui() -> gr.Blocks:
|
|
400 |
# ==================================================================================
|
401 |
|
402 |
if __name__ == "__main__":
|
403 |
-
# Ensure cache directory exists
|
404 |
os.makedirs(CACHE_DIR, exist_ok=True)
|
405 |
-
|
406 |
app = create_ui()
|
407 |
-
app.queue().launch(
|
|
|
12 |
import onnxruntime as rt
|
13 |
from PIL import Image
|
14 |
from huggingface_hub import hf_hub_download
|
15 |
+
from transformers import pipeline, Pipeline, AutoModel, AutoProcessor
|
16 |
from tqdm import tqdm
|
17 |
|
18 |
+
# Suppress a specific PIL warning about image size to handle large images
|
19 |
Image.MAX_IMAGE_PIXELS = None
|
20 |
|
21 |
# --- Configuration ---
|
22 |
CACHE_DIR = "./hf_cache"
|
23 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
24 |
+
# Use bfloat16 for modern GPUs, float32 for others (including CPU)
|
25 |
DTYPE = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8 else torch.float32
|
26 |
|
27 |
print(f"Using device: {DEVICE} with dtype: {DTYPE}")
|
|
|
32 |
|
33 |
class AestheticScorer(ABC):
|
34 |
"""Abstract base class for all aesthetic scoring models."""
|
|
|
35 |
def __init__(self, model_name: str, repo_id: str, filename: str = None):
|
36 |
self.model_name = model_name
|
37 |
self.repo_id = repo_id
|
38 |
self.filename = filename
|
39 |
self._model = None
|
40 |
+
print(f"Initializing scorer definition: {self.model_name}")
|
41 |
|
42 |
@property
|
43 |
def model(self):
|
44 |
"""Lazy-loads the model on first access."""
|
45 |
if self._model is None:
|
46 |
+
print(f"Loading model weights for '{self.model_name}'...")
|
47 |
self._model = self.load_model()
|
48 |
+
print(f"'{self.model_name}' model weights loaded.")
|
49 |
return self._model
|
50 |
|
51 |
def _download_model(self) -> str:
|
|
|
63 |
pass
|
64 |
|
65 |
def release_model(self):
|
66 |
+
"""Releases model from memory to conserve VRAM/RAM."""
|
67 |
if self._model is not None:
|
68 |
+
print(f"Releasing model from memory: {self.model_name}")
|
69 |
del self._model
|
70 |
self._model = None
|
71 |
gc.collect()
|
|
|
74 |
|
75 |
class PipelineScorer(AestheticScorer):
|
76 |
"""Scorer for models compatible with Hugging Face pipelines."""
|
|
|
77 |
def load_model(self) -> Pipeline:
|
78 |
+
return pipeline("image-classification", model=self.repo_id, device=DEVICE)
|
|
|
|
|
|
|
|
|
|
|
79 |
|
80 |
@torch.no_grad()
|
81 |
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
82 |
+
results = self.model(image_batch, top_k=None) # Get all class scores
|
|
|
83 |
scores = []
|
84 |
for res in results:
|
85 |
try:
|
|
|
86 |
hq_score = next(item['score'] for item in res if item['label'] == 'hq')
|
87 |
scores.append(round(hq_score * 10.0, 4))
|
88 |
except (StopIteration, TypeError):
|
|
|
91 |
|
92 |
class ONNXScorer(AestheticScorer):
|
93 |
"""Scorer for ONNX-based models."""
|
|
|
94 |
def load_model(self) -> rt.InferenceSession:
|
|
|
95 |
model_path = self._download_model()
|
96 |
return rt.InferenceSession(model_path, providers=['CUDAExecutionProvider' if DEVICE == 'cuda' else 'CPUExecutionProvider'])
|
97 |
|
98 |
def _preprocess(self, img: Image.Image) -> np.ndarray:
|
|
|
99 |
img_np = np.array(img.convert("RGB")).astype(np.float32) / 255.0
|
100 |
s = 768
|
101 |
h, w = img_np.shape[:2]
|
102 |
+
ratio = s / max(h, w)
|
103 |
+
new_h, new_w = int(h * ratio), int(w * ratio)
|
|
|
|
|
104 |
|
105 |
resized = cv2.resize(img_np, (new_w, new_h), interpolation=cv2.INTER_AREA)
|
106 |
canvas = np.zeros((s, s, 3), dtype=np.float32)
|
|
|
110 |
return np.transpose(canvas, (2, 0, 1))[np.newaxis, :]
|
111 |
|
112 |
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
|
|
113 |
scores = []
|
114 |
for img in image_batch:
|
115 |
try:
|
|
|
121 |
return scores
|
122 |
|
123 |
class CLIPMLPScorer(AestheticScorer):
|
124 |
+
"""Scorer for models using a CLIP backbone and a custom MLP head."""
|
|
|
125 |
class MLP(torch.nn.Module):
|
126 |
+
"""Re-implementation of the exact MLP from the original code."""
|
127 |
def __init__(self, input_size: int):
|
128 |
super().__init__()
|
129 |
self.layers = torch.nn.Sequential(
|
130 |
+
torch.nn.Linear(input_size, 2048),
|
131 |
torch.nn.ReLU(),
|
132 |
+
torch.nn.BatchNorm1d(2048),
|
133 |
+
torch.nn.Dropout(0.3),
|
134 |
+
torch.nn.Linear(2048, 512),
|
135 |
+
torch.nn.ReLU(),
|
136 |
+
torch.nn.BatchNorm1d(512),
|
137 |
+
torch.nn.Dropout(0.3),
|
138 |
+
torch.nn.Linear(512, 256),
|
139 |
torch.nn.ReLU(),
|
140 |
+
torch.nn.BatchNorm1d(256),
|
141 |
torch.nn.Dropout(0.2),
|
142 |
+
torch.nn.Linear(256, 128),
|
143 |
torch.nn.ReLU(),
|
144 |
+
torch.nn.BatchNorm1d(128),
|
145 |
+
torch.nn.Dropout(0.1),
|
146 |
+
torch.nn.Linear(128, 32),
|
147 |
torch.nn.ReLU(),
|
148 |
+
torch.nn.Linear(32, 1)
|
149 |
)
|
150 |
def forward(self, x):
|
151 |
return self.layers(x)
|
152 |
|
153 |
def load_model(self) -> Dict[str, Any]:
|
154 |
+
import clip
|
|
|
|
|
155 |
model_path = self._download_model()
|
|
|
156 |
mlp = self.MLP(input_size=768) # ViT-L/14 has 768 features
|
157 |
state_dict = torch.load(model_path, map_location=DEVICE)
|
158 |
mlp.load_state_dict(state_dict)
|
159 |
+
mlp.to(device=DEVICE)
|
160 |
mlp.eval()
|
|
|
161 |
clip_model, preprocess = clip.load("ViT-L/14", device=DEVICE)
|
|
|
162 |
return {"mlp": mlp, "clip": clip_model, "preprocess": preprocess}
|
163 |
|
164 |
@torch.no_grad()
|
165 |
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
|
|
166 |
preprocess = self.model['preprocess']
|
167 |
+
# Handle single-image batches correctly for CLIP
|
168 |
+
if len(image_batch) == 1:
|
169 |
+
image_batch = image_batch * 2
|
170 |
+
single_image_mode = True
|
171 |
+
else:
|
172 |
+
single_image_mode = False
|
173 |
+
|
174 |
image_tensors = torch.cat([preprocess(img).unsqueeze(0) for img in image_batch]).to(DEVICE)
|
175 |
+
image_features = self.model['clip'].encode_image(image_tensors).to(torch.float32)
|
|
|
176 |
image_features /= image_features.norm(dim=-1, keepdim=True)
|
177 |
+
predictions = self.model['mlp'](image_features).squeeze(-1)
|
178 |
+
scores = predictions.clamp(0, 10).float().cpu().numpy()
|
179 |
|
180 |
+
final_scores = [round(float(s), 4) for s in scores]
|
181 |
+
return final_scores[:1] if single_image_mode else final_scores
|
182 |
+
|
183 |
+
class SigLIPScorer(AestheticScorer):
|
184 |
+
"""Scorer for the Aesthetic Predictor V2.5 SigLIP model."""
|
185 |
+
def load_model(self) -> Dict[str, Any]:
|
186 |
+
model = AutoModel.from_pretrained(self.repo_id, trust_remote_code=True).to(DEVICE, DTYPE).eval()
|
187 |
+
processor = AutoProcessor.from_pretrained(self.repo_id, trust_remote_code=True)
|
188 |
+
return {"model": model, "processor": processor}
|
189 |
+
|
190 |
+
@torch.no_grad()
|
191 |
+
def score_batch(self, image_batch: List[Image.Image]) -> List[float]:
|
192 |
+
inputs = self.model['processor'](
|
193 |
+
images=[img.convert("RGB") for img in image_batch],
|
194 |
+
return_tensors="pt"
|
195 |
+
)
|
196 |
+
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
|
197 |
+
inputs['pixel_values'] = inputs['pixel_values'].to(DTYPE)
|
198 |
+
logits = self.model(**inputs).logits.squeeze(-1)
|
199 |
+
scores = logits.float().cpu().numpy()
|
200 |
return [round(float(s), 4) for s in scores]
|
201 |
|
202 |
# --- Model Registry ---
|
203 |
+
MODEL_REGISTRY: Dict[str, AestheticScorer] = {
|
204 |
+
"Aesthetic Shadow V2": PipelineScorer("Aesthetic Shadow V2", "NeoChen1024/aesthetic-shadow-v2-backup"),
|
205 |
+
"Waifu Scorer V3": CLIPMLPScorer("Waifu Scorer V3", "Eugeoter/waifu-scorer-v3", "model.pth"),
|
206 |
+
"Aesthetic V2.5 SigLIP": SigLIPScorer("Aesthetic V2.5 SigLIP", "জিংוניत्र/Aesthetic-Predictor-V2-5-SigLIP"),
|
207 |
+
"Anime Scorer": ONNXScorer("Anime Scorer", "skytnt/anime-aesthetic", "model.onnx")
|
|
|
|
|
|
|
|
|
|
|
208 |
}
|
|
|
|
|
209 |
_loaded_models_cache: Dict[str, AestheticScorer] = {}
|
210 |
|
|
|
211 |
# ==================================================================================
|
212 |
# 2. CORE PROCESSING LOGIC
|
213 |
# ==================================================================================
|
214 |
|
215 |
def get_scorers(model_names: List[str]) -> List[AestheticScorer]:
|
216 |
"""Retrieves and caches scorer instances based on selected names."""
|
217 |
+
for name in list(_loaded_models_cache.keys()):
|
|
|
218 |
if name not in model_names:
|
219 |
+
_loaded_models_cache[name].release_model()
|
220 |
del _loaded_models_cache[name]
|
221 |
+
return [_loaded_models_cache.setdefault(name, MODEL_REGISTRY[name]) for name in model_names]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
|
223 |
def evaluate_images(
|
224 |
+
files: List[gr.File], selected_model_names: List[str], batch_size: int, progress=gr.Progress(track_tqdm=True)
|
|
|
|
|
|
|
225 |
) -> pd.DataFrame:
|
226 |
+
"""Main function to process images and return results as a Pandas DataFrame."""
|
|
|
|
|
|
|
227 |
if not files:
|
228 |
gr.Warning("No images uploaded. Please upload files to evaluate.")
|
229 |
return pd.DataFrame()
|
|
|
230 |
if not selected_model_names:
|
231 |
gr.Warning("No models selected. Please select at least one model.")
|
232 |
return pd.DataFrame()
|
233 |
|
234 |
try:
|
235 |
image_paths = [Path(f.name) for f in files]
|
236 |
+
all_results, scorers = [], get_scorers(selected_model_names)
|
|
|
|
|
|
|
|
|
237 |
|
238 |
+
for i in tqdm(range(0, len(image_paths), batch_size), desc="Processing Batches"):
|
239 |
batch_paths = image_paths[i : i + batch_size]
|
|
|
|
|
240 |
try:
|
241 |
batch_images = [Image.open(p).convert("RGB") for p in batch_paths]
|
242 |
except Exception as e:
|
243 |
gr.Warning(f"Skipping a batch due to an error loading an image: {e}")
|
|
|
244 |
continue
|
|
|
|
|
|
|
|
|
|
|
245 |
|
246 |
+
batch_scores = {scorer.model_name: scorer.score_batch(batch_images) for scorer in scorers}
|
247 |
+
|
248 |
for j, path in enumerate(batch_paths):
|
249 |
+
result_row = {"Image": str(path), "Filename": path.name}
|
250 |
+
scores_for_avg = [batch_scores[s.model_name][j] for s in scorers]
|
|
|
251 |
for scorer in scorers:
|
252 |
+
result_row[scorer.model_name] = batch_scores[scorer.model_name][j]
|
253 |
+
result_row["Average Score"] = round(np.mean(scores_for_avg), 4) if scores_for_avg else 0.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
all_results.append(result_row)
|
|
|
|
|
255 |
|
256 |
+
return pd.DataFrame(all_results) if all_results else pd.DataFrame()
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
|
258 |
except Exception as e:
|
259 |
gr.Error(f"A critical error occurred: {e}")
|
|
|
|
|
|
|
|
|
260 |
return pd.DataFrame()
|
261 |
|
|
|
262 |
# ==================================================================================
|
263 |
# 3. GRADIO USER INTERFACE
|
264 |
# ==================================================================================
|
265 |
|
266 |
def create_ui() -> gr.Blocks:
|
267 |
"""Creates and configures the Gradio web interface."""
|
|
|
268 |
all_model_names = list(MODEL_REGISTRY.keys())
|
|
|
|
|
269 |
dataframe_headers = ["Image", "Filename"] + all_model_names + ["Average Score"]
|
270 |
dataframe_datatypes = ["image", "str"] + ["number"] * (len(all_model_names) + 1)
|
271 |
|
272 |
+
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), title="Image Aesthetic Scorer") as demo:
|
273 |
+
gr.Markdown("# 🖼️ Modern Image Aesthetic Scorer")
|
274 |
+
gr.Markdown("Upload images, select models, and click 'Evaluate'. Results table supports **interactive sorting** and **downloading as CSV**.")
|
|
|
|
|
|
|
|
|
|
|
275 |
|
276 |
with gr.Row():
|
277 |
with gr.Column(scale=1):
|
278 |
+
input_files = gr.Files(label="Upload Images", file_count="multiple", file_types=["image"])
|
279 |
+
model_checkboxes = gr.CheckboxGroup(choices=all_model_names, value=all_model_names, label="Scoring Models")
|
280 |
+
batch_size_slider = gr.Slider(minimum=1, maximum=64, value=8, step=1, label="Batch Size", info="Adjust based on your VRAM.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
281 |
with gr.Row():
|
282 |
process_button = gr.Button("🚀 Evaluate Images", variant="primary")
|
283 |
clear_button = gr.Button("🧹 Clear All")
|
284 |
|
285 |
with gr.Column(scale=3):
|
286 |
+
# CORRECTED LINE: height and show_download_button are passed directly here.
|
287 |
results_dataframe = gr.DataFrame(
|
288 |
headers=dataframe_headers,
|
289 |
datatype=dataframe_datatypes,
|
290 |
label="Evaluation Scores",
|
291 |
interactive=True,
|
292 |
+
height=800,
|
293 |
+
show_download_button=True
|
294 |
)
|
|
|
|
|
|
|
295 |
|
|
|
296 |
process_button.click(
|
297 |
fn=evaluate_images,
|
298 |
inputs=[input_files, model_checkboxes, batch_size_slider],
|
299 |
+
outputs=[results_dataframe]
|
|
|
300 |
)
|
301 |
|
302 |
def clear_outputs():
|
303 |
+
for scorer in list(_loaded_models_cache.values()):
|
|
|
304 |
scorer.release_model()
|
305 |
_loaded_models_cache.clear()
|
306 |
gr.Info("Cleared results and released models from memory.")
|
307 |
+
return pd.DataFrame(), None # Clear dataframe and file input
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
308 |
|
309 |
+
clear_button.click(fn=clear_outputs, outputs=[results_dataframe, input_files])
|
310 |
return demo
|
311 |
|
312 |
# ==================================================================================
|
|
|
314 |
# ==================================================================================
|
315 |
|
316 |
if __name__ == "__main__":
|
|
|
317 |
os.makedirs(CACHE_DIR, exist_ok=True)
|
|
|
318 |
app = create_ui()
|
319 |
+
app.queue().launch()
|