File size: 16,212 Bytes
8debb94
 
4477181
 
 
 
 
8debb94
 
4477181
 
8debb94
161be00
38783e1
8debb94
 
 
 
4477181
 
161be00
8debb94
38783e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4477181
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8debb94
 
4477181
 
8debb94
512487a
 
8debb94
 
161be00
4477181
 
 
512487a
4477181
8debb94
 
5ac51d9
4477181
8debb94
4477181
 
 
512487a
4477181
8debb94
 
5ac51d9
4477181
8debb94
4477181
 
 
512487a
4477181
8debb94
 
5ac51d9
4477181
8debb94
4477181
38783e1
 
4477181
512487a
4477181
 
8debb94
 
4477181
8debb94
 
4477181
38783e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4477181
 
 
 
8debb94
 
 
 
 
4477181
8debb94
 
 
 
 
4477181
 
8debb94
512487a
 
 
8debb94
 
 
 
 
 
4477181
 
 
8debb94
4477181
 
161be00
8debb94
 
4477181
38783e1
 
4477181
161be00
4477181
 
 
 
38783e1
4477181
161be00
4477181
 
 
38783e1
 
 
 
 
 
4477181
161be00
 
4477181
 
161be00
 
4477181
 
8debb94
38783e1
8debb94
4477181
38783e1
4477181
 
38783e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4477181
38783e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8debb94
 
 
 
4477181
38783e1
8debb94
 
 
38783e1
 
8debb94
 
 
 
 
4477181
8debb94
 
 
 
38783e1
 
 
 
 
 
 
 
 
8debb94
4477181
8debb94
4477181
38783e1
 
 
 
 
4477181
38783e1
8debb94
38783e1
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
#!/usr/bin/env python3
"""
Example usage of GAP-CLIP models.

This file provides example code for loading and using the models (color,
hierarchy, main) from local checkpoints or the Hugging Face Hub. It shows
how to load models, extract embeddings, and perform similarity comparisons.
"""

import os

import torch
import torch.nn.functional as F
import requests
from PIL import Image
from transformers import CLIPProcessor, CLIPModel as CLIPModel_transformers
from huggingface_hub import hf_hub_download

from training.color_model import ColorCLIP
from training.hierarchy_model import HierarchyModel
import config

CLIP_MODEL_NAME = "laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
HF_REPO_ID = "Leacb4/gap-clip"


# ---------------------------------------------------------------------------
# Simple API — load from HF and get 512D embeddings
# ---------------------------------------------------------------------------

def load_gap_clip(repo_id: str = HF_REPO_ID):
    """
    Load the GAP-CLIP model directly from Hugging Face.

    This is the simplest way to use the model. Returns (model, processor).

    Example::

        model, processor = load_gap_clip()
        emb = get_image_embedding_from_url(
            "https://www.gap.com/webcontent/0060/662/817/cn60662817.jpg",
            model, processor,
        )
        print(emb.shape)  # torch.Size([1, 512])
    """
    model = CLIPModel_transformers.from_pretrained(repo_id)
    processor = CLIPProcessor.from_pretrained(repo_id)
    model.eval()
    return model, processor


def get_image_embedding_from_url(url: str, model, processor, device=None):
    """
    Download an image from a URL and return its 512D GAP-CLIP embedding.

    Args:
        url: Image URL.
        model: CLIPModel loaded via load_gap_clip() or from_pretrained().
        processor: CLIPProcessor matching the model.
        device: Device to run on (defaults to config.device).

    Returns:
        Tensor of shape [1, 512] (L2-normalized).
    """
    device = device or config.device
    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
    inputs = processor(images=image, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    model = model.to(device)
    with torch.no_grad():
        image_features = model.get_image_features(**inputs)
    return F.normalize(image_features, dim=-1)


def get_text_embedding(text: str, model, processor, device=None):
    """
    Return a 512D GAP-CLIP embedding for a text query.

    Args:
        text: Text query (e.g., "red dress").
        model: CLIPModel loaded via load_gap_clip() or from_pretrained().
        processor: CLIPProcessor matching the model.
        device: Device to run on (defaults to config.device).

    Returns:
        Tensor of shape [1, 512] (L2-normalized).
    """
    device = device or config.device
    inputs = processor(text=[text], return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    model = model.to(device)
    with torch.no_grad():
        text_features = model.get_text_features(**inputs)
    return F.normalize(text_features, dim=-1)


# ---------------------------------------------------------------------------
# Internal helpers for encode_text / encode_image (used by advanced examples)
# ---------------------------------------------------------------------------

def encode_text(model, processor, text_queries, device):
    """Encode text queries into embeddings (unnormalized)."""
    if isinstance(text_queries, str):
        text_queries = [text_queries]
    inputs = processor(text=text_queries, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    with torch.no_grad():
        text_features = model.get_text_features(**inputs)
    return text_features


def encode_image(model, processor, images, device):
    """Encode images into embeddings (unnormalized)."""
    if not isinstance(images, list):
        images = [images]
    inputs = processor(images=images, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    with torch.no_grad():
        image_features = model.get_image_features(**inputs)
    return image_features


def load_models_from_hf(repo_id: str, cache_dir: str = "./models_cache"):
    """
    Load models from Hugging Face.

    Args:
        repo_id: ID of the Hugging Face repository
        cache_dir: Local cache directory
    """
    os.makedirs(cache_dir, exist_ok=True)
    device = config.device

    print(f"Loading models from '{repo_id}'...")

    # 1. Loading color model
    print("  Loading color model...")
    color_model_path = hf_hub_download(
        repo_id=repo_id,
        filename="models/color_model.pt",
        cache_dir=cache_dir,
    )
    color_model = ColorCLIP.from_checkpoint(color_model_path, device=device)
    print("  Color model loaded")

    # 2. Loading hierarchy model
    print("  Loading hierarchy model...")
    hierarchy_model_path = hf_hub_download(
        repo_id=repo_id,
        filename="models/hierarchy_model.pth",
        cache_dir=cache_dir,
    )
    hierarchy_model = HierarchyModel.from_checkpoint(hierarchy_model_path, device=device)
    print("  Hierarchy model loaded")

    # 3. Loading main CLIP model
    print("  Loading main CLIP model...")
    main_model_path = hf_hub_download(
        repo_id=repo_id,
        filename="models/gap_clip.pth",
        cache_dir=cache_dir,
    )

    clip_model = CLIPModel_transformers.from_pretrained(CLIP_MODEL_NAME)
    checkpoint = torch.load(main_model_path, map_location=device, weights_only=False)

    # Handle different checkpoint structures
    if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint:
        clip_model.load_state_dict(checkpoint['model_state_dict'])
    else:
        clip_model.load_state_dict(checkpoint)

    clip_model = clip_model.to(device)
    clip_model.eval()

    processor = CLIPProcessor.from_pretrained(CLIP_MODEL_NAME)
    print("  Main CLIP model loaded")

    print("\nAll models loaded!")

    return {
        'color_model': color_model,
        'hierarchy_model': hierarchy_model,
        'main_model': clip_model,
        'processor': processor,
        'device': device,
    }


def load_models_from_local(
    color_model_path: str = None,
    hierarchy_model_path: str = None,
    main_model_path: str = None,
):
    """
    Load models from local checkpoint files.

    Args:
        color_model_path: Path to color_model.pt (defaults to config.color_model_path)
        hierarchy_model_path: Path to hierarchy_model.pth (defaults to config.hierarchy_model_path)
        main_model_path: Path to gap_clip.pth (defaults to config.main_model_path)
    """
    device = config.device
    color_model_path = color_model_path or config.color_model_path
    hierarchy_model_path = hierarchy_model_path or config.hierarchy_model_path
    main_model_path = main_model_path or config.main_model_path

    print(f"Loading models from local checkpoints (device={device})...")

    # 1. Color model
    print("  Loading color model...")
    color_model = ColorCLIP.from_checkpoint(color_model_path, device=device)
    print("  Color model loaded")

    # 2. Hierarchy model
    print("  Loading hierarchy model...")
    hierarchy_model = HierarchyModel.from_checkpoint(hierarchy_model_path, device=device)
    print("  Hierarchy model loaded")

    # 3. Main CLIP model
    print("  Loading main CLIP model...")
    clip_model = CLIPModel_transformers.from_pretrained(CLIP_MODEL_NAME)
    checkpoint = torch.load(main_model_path, map_location=device, weights_only=False)
    if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint:
        clip_model.load_state_dict(checkpoint['model_state_dict'])
    else:
        clip_model.load_state_dict(checkpoint)
    clip_model.to(device).eval()
    processor = CLIPProcessor.from_pretrained(CLIP_MODEL_NAME)
    print("  Main CLIP model loaded")

    print("\nAll models loaded!")

    return {
        'color_model': color_model,
        'hierarchy_model': hierarchy_model,
        'main_model': clip_model,
        'processor': processor,
        'device': device,
    }


def example_search(models, image_path: str = None, text_query: str = None):
    """
    Example search with the models.

    Args:
        models: Dictionary of loaded models
        image_path: Path to an image (optional)
        text_query: Text query (optional)
    """
    color_model = models['color_model']
    hierarchy_model = models['hierarchy_model']
    main_model = models['main_model']
    processor = models['processor']
    device = models['device']

    print("\nExample search...")

    if text_query:
        print(f"  Text query: '{text_query}'")

        # Get color and hierarchy embeddings
        color_emb = color_model.get_text_embeddings([text_query])
        hierarchy_emb = hierarchy_model.get_text_embeddings([text_query])

        print(f"  Color embedding shape: {color_emb.shape}, norm: {color_emb.norm(dim=-1).item():.4f}")
        print(f"  Hierarchy embedding shape: {hierarchy_emb.shape}, norm: {hierarchy_emb.norm(dim=-1).item():.4f}")

        # Get main model embeddings
        text_features = encode_text(main_model, processor, text_query, device)
        text_features = F.normalize(text_features, dim=-1)

        print(f"  Main embedding: {text_features.shape}")
        print(f"  First 10 dims of main embedding: {text_features[0, :10]}")

        # Extract color and hierarchy embeddings from main embedding
        main_color_emb = text_features[:, :config.color_emb_dim]
        main_hierarchy_emb = text_features[:, config.color_emb_dim:config.color_emb_dim + config.hierarchy_emb_dim]

        print(f"\n  Subspace comparison (color model vs main model dims [0:{config.color_emb_dim}]):")
        print(f"    color_model first 5 dims: {color_emb[0, :5].tolist()}")
        print(f"    main_model  first 5 dims: {main_color_emb[0, :5].tolist()}")
        print(f"  Subspace comparison (hierarchy model vs main model dims [{config.color_emb_dim}:{config.color_emb_dim + config.hierarchy_emb_dim}]):")
        print(f"    hierarchy_model first 5 dims: {hierarchy_emb[0, :5].tolist()}")
        print(f"    main_model      first 5 dims: {main_hierarchy_emb[0, :5].tolist()}")

        # Calculate cosine similarity between color embeddings
        color_cosine_sim = F.cosine_similarity(color_emb, main_color_emb, dim=1)
        print(f"\n  Cosine similarity between color embeddings: {color_cosine_sim.item():.4f}")

        # Calculate cosine similarity between hierarchy embeddings
        hierarchy_cosine_sim = F.cosine_similarity(hierarchy_emb, main_hierarchy_emb, dim=1)
        print(f"  Cosine similarity between hierarchy embeddings: {hierarchy_cosine_sim.item():.4f}")

    if image_path and os.path.exists(image_path):
        print(f"\n  Image: {image_path}")
        image = Image.open(image_path).convert("RGB")

        # Main model image embedding
        image_features = encode_image(main_model, processor, image, device)
        image_features = F.normalize(image_features, dim=-1)
        print(f"  Main image embedding shape: {image_features.shape}")

        # Color model image embedding (preprocess through model's own processor)
        color_pixel_values = color_model.processor(
            images=image, return_tensors="pt"
        )["pixel_values"].to(device)
        color_img_emb = color_model.get_image_embeddings(color_pixel_values)
        print(f"  Color image embedding shape: {color_img_emb.shape}")

        # Hierarchy model image embedding
        hierarchy_pixel_values = hierarchy_model.processor(
            images=image, return_tensors="pt"
        )["pixel_values"].to(device)
        hierarchy_img_emb = hierarchy_model.get_image_embeddings(hierarchy_pixel_values)
        print(f"  Hierarchy image embedding shape: {hierarchy_img_emb.shape}")

        # Compare subspace alignment for images
        main_color_img = image_features[:, :config.color_emb_dim]
        main_hierarchy_img = image_features[:, config.color_emb_dim:config.color_emb_dim + config.hierarchy_emb_dim]
        color_img_sim = F.cosine_similarity(color_img_emb, main_color_img, dim=1)
        hierarchy_img_sim = F.cosine_similarity(hierarchy_img_emb, main_hierarchy_img, dim=1)
        print(f"  Image color subspace cosine similarity: {color_img_sim.item():.4f}")
        print(f"  Image hierarchy subspace cosine similarity: {hierarchy_img_sim.item():.4f}")


def example_similarity_search(models, image_paths: list, text_query: str):
    """
    Rank images by similarity to a text query using GAP-CLIP.

    Shows the key use case: computing text-to-image similarity scores
    for ranking, combining color, hierarchy, and general CLIP subspaces.

    Args:
        models: Dictionary of loaded models
        image_paths: List of image file paths to rank
        text_query: Text query to match against
    """
    main_model = models['main_model']
    processor = models['processor']
    device = models['device']

    print(f"\nSimilarity search: '{text_query}' against {len(image_paths)} images")

    # Encode the text query
    text_features = encode_text(main_model, processor, text_query, device)
    text_features = F.normalize(text_features, dim=-1)  # [1, 512]

    # Encode all images
    images = []
    valid_paths = []
    for p in image_paths:
        if os.path.exists(p):
            images.append(Image.open(p).convert("RGB"))
            valid_paths.append(p)
        else:
            print(f"  Warning: {p} not found, skipping")

    if not images:
        print("  No valid images found.")
        return

    image_features = encode_image(main_model, processor, images, device)
    image_features = F.normalize(image_features, dim=-1)  # [N, 512]

    # Full 512D similarity
    full_scores = (text_features @ image_features.T).squeeze(0)  # [N]

    # Subspace similarities
    color_dim = config.color_emb_dim
    hierarchy_end = color_dim + config.hierarchy_emb_dim

    color_text = F.normalize(text_features[:, :color_dim], dim=-1)
    color_imgs = F.normalize(image_features[:, :color_dim], dim=-1)
    color_scores = (color_text @ color_imgs.T).squeeze(0)

    hier_text = F.normalize(text_features[:, color_dim:hierarchy_end], dim=-1)
    hier_imgs = F.normalize(image_features[:, color_dim:hierarchy_end], dim=-1)
    hierarchy_scores = (hier_text @ hier_imgs.T).squeeze(0)

    # Rank by full similarity
    ranked_indices = full_scores.argsort(descending=True)

    print(f"\n  Ranking (by full 512D cosine similarity):")
    for rank, idx in enumerate(ranked_indices):
        i = idx.item()
        print(
            f"    {rank + 1}. {os.path.basename(valid_paths[i]):30s}"
            f"  full={full_scores[i]:.4f}"
            f"  color={color_scores[i]:.4f}"
            f"  hierarchy={hierarchy_scores[i]:.4f}"
        )


if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="Example usage of GAP-CLIP models")
    parser.add_argument(
        "--repo-id",
        type=str,
        default=None,
        help="Hugging Face repo ID (e.g., Leacb4/gap-clip). If omitted, loads from local paths.",
    )
    parser.add_argument(
        "--text",
        type=str,
        default="red dress",
        help="Text query for search",
    )
    parser.add_argument(
        "--image",
        type=str,
        default=None,
        help="Path to a single image for example_search",
    )
    parser.add_argument(
        "--images",
        type=str,
        nargs="+",
        default=None,
        help="Paths to multiple images for similarity ranking",
    )

    args = parser.parse_args()

    # Load models (HF or local)
    if args.repo_id:
        models = load_models_from_hf(args.repo_id)
    else:
        models = load_models_from_local()

    # Example search (embedding inspection)
    example_search(models, image_path=args.image, text_query=args.text)

    # Similarity ranking (if multiple images provided)
    if args.images:
        example_similarity_search(models, args.images, args.text)