File size: 12,672 Bytes
86f402d
 
 
 
 
1a97904
86f402d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1a97904
 
 
 
 
86f402d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
MONET Tool - Skin lesion feature extraction using MONET model
Correct implementation based on MONET tutorial: automatic_concept_annotation.ipynb
"""

import os
import torch
import torch.nn.functional as F
import numpy as np
import scipy.special
from PIL import Image
from typing import Optional, Dict, List
import torchvision.transforms as T


# The 7 MONET feature columns expected by ConvNeXt
MONET_FEATURES = [
    "MONET_ulceration_crust",
    "MONET_hair",
    "MONET_vasculature_vessels",
    "MONET_erythema",
    "MONET_pigmented",
    "MONET_gel_water_drop_fluid_dermoscopy_liquid",
    "MONET_skin_markings_pen_ink_purple_pen",
]

# Concept terms for each MONET feature (multiple synonyms improve detection)
MONET_CONCEPT_TERMS = {
    "MONET_ulceration_crust": ["ulceration", "crust", "crusting", "ulcer"],
    "MONET_hair": ["hair", "hairy"],
    "MONET_vasculature_vessels": ["blood vessels", "vasculature", "vessels", "telangiectasia"],
    "MONET_erythema": ["erythema", "redness", "red"],
    "MONET_pigmented": ["pigmented", "pigmentation", "melanin", "brown"],
    "MONET_gel_water_drop_fluid_dermoscopy_liquid": ["dermoscopy gel", "fluid", "water drop", "immersion fluid"],
    "MONET_skin_markings_pen_ink_purple_pen": ["pen marking", "ink", "surgical marking", "purple pen"],
}

# Prompt templates (from MONET paper)
PROMPT_TEMPLATES = [
    "This is skin image of {}",
    "This is dermatology image of {}",
    "This is image of {}",
]

# Reference prompts (baseline for contrastive scoring)
PROMPT_REFS = [
    ["This is skin image"],
    ["This is dermatology image"],
    ["This is image"],
]


def get_transform(n_px=224):
    """Get MONET preprocessing transform"""
    def convert_image_to_rgb(image):
        return image.convert("RGB")

    return T.Compose([
        T.Resize(n_px, interpolation=T.InterpolationMode.BICUBIC),
        T.CenterCrop(n_px),
        convert_image_to_rgb,
        T.ToTensor(),
        T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
    ])


class MonetTool:
    """
    MONET tool for extracting concept presence scores from skin lesion images.
    Uses the proper contrastive scoring method from the MONET paper.
    """

    def __init__(self, device: Optional[str] = None, use_hf: bool = True):
        """
        Args:
            device: Device to run on (cuda, mps, cpu)
            use_hf: Use HuggingFace implementation (True) or original CLIP (False)
        """
        self.model = None
        self.processor = None
        self.device = device
        self.use_hf = use_hf
        self.loaded = False
        self.transform = get_transform(224)

        # Cache for concept embeddings
        self._concept_embeddings = {}

    def load(self):
        """Load MONET model"""
        if self.loaded:
            return

        # Determine device (respect SKINPRO_TOOL_DEVICE override for GPU sharing)
        forced = os.environ.get("SKINPRO_TOOL_DEVICE")
        if forced:
            self.device = forced
        elif self.device is None:
            if torch.cuda.is_available():
                self.device = "cuda:0"
            elif torch.backends.mps.is_available():
                self.device = "mps"
            else:
                self.device = "cpu"

        if self.use_hf:
            # HuggingFace implementation
            from transformers import AutoProcessor, AutoModelForZeroShotImageClassification

            self.processor = AutoProcessor.from_pretrained("chanwkim/monet")
            self.model = AutoModelForZeroShotImageClassification.from_pretrained("chanwkim/monet")
            self.model.to(self.device)
            self.model.eval()
        else:
            # Original CLIP implementation
            import clip

            self.model, _ = clip.load("ViT-L/14", device=self.device, jit=False)
            self.model.load_state_dict(
                torch.hub.load_state_dict_from_url(
                    "https://aimslab.cs.washington.edu/MONET/weight_clip.pt"
                )
            )
            self.model.eval()

        self.loaded = True

        # Pre-compute concept embeddings for all MONET features
        self._precompute_concept_embeddings()

    def _encode_text(self, text_list: List[str]) -> torch.Tensor:
        """Encode text to embeddings"""
        if self.use_hf:
            inputs = self.processor(text=text_list, return_tensors="pt", padding=True)
            inputs = {k: v.to(self.device) for k, v in inputs.items()}
            with torch.no_grad():
                embeddings = self.model.get_text_features(**inputs)
        else:
            import clip
            tokens = clip.tokenize(text_list, truncate=True).to(self.device)
            with torch.no_grad():
                embeddings = self.model.encode_text(tokens)

        return embeddings.cpu()

    def _encode_image(self, image: Image.Image) -> torch.Tensor:
        """Encode image to embedding"""
        image_tensor = self.transform(image).unsqueeze(0).to(self.device)

        if self.use_hf:
            with torch.no_grad():
                embedding = self.model.get_image_features(image_tensor)
        else:
            with torch.no_grad():
                embedding = self.model.encode_image(image_tensor)

        return embedding.cpu()

    def _precompute_concept_embeddings(self):
        """Pre-compute embeddings for all MONET concepts"""
        for feature_name, concept_terms in MONET_CONCEPT_TERMS.items():
            self._concept_embeddings[feature_name] = self._get_concept_embedding(concept_terms)

    def _get_concept_embedding(self, concept_terms: List[str]) -> Dict:
        """
        Generate prompt embeddings for a concept using multiple templates.

        Args:
            concept_terms: List of synonymous terms for the concept

        Returns:
            dict with target and reference embeddings
        """
        # Target prompts: "This is skin image of {term}"
        prompt_target = [
            [template.format(term) for term in concept_terms]
            for template in PROMPT_TEMPLATES
        ]

        # Flatten and encode
        prompt_target_flat = [p for template_prompts in prompt_target for p in template_prompts]
        target_embeddings = self._encode_text(prompt_target_flat)

        # Reshape to [num_templates, num_terms, embed_dim]
        num_templates = len(PROMPT_TEMPLATES)
        num_terms = len(concept_terms)
        embed_dim = target_embeddings.shape[-1]
        target_embeddings = target_embeddings.view(num_templates, num_terms, embed_dim)

        # Normalize
        target_embeddings_norm = F.normalize(target_embeddings, dim=2)

        # Reference prompts: "This is skin image"
        prompt_ref_flat = [p for ref_list in PROMPT_REFS for p in ref_list]
        ref_embeddings = self._encode_text(prompt_ref_flat)
        ref_embeddings = ref_embeddings.view(num_templates, -1, embed_dim)
        ref_embeddings_norm = F.normalize(ref_embeddings, dim=2)

        return {
            "target_embedding_norm": target_embeddings_norm,
            "ref_embedding_norm": ref_embeddings_norm,
        }

    def _calculate_concept_score(
        self,
        image_features_norm: torch.Tensor,
        concept_embedding: Dict,
        temp: float = 1 / np.exp(4.5944)
    ) -> float:
        """
        Calculate concept presence score using contrastive comparison.

        Args:
            image_features_norm: Normalized image embedding [1, embed_dim]
            concept_embedding: Dict with target and reference embeddings
            temp: Temperature for softmax

        Returns:
            Concept presence score (0-1)
        """
        target_emb = concept_embedding["target_embedding_norm"].float()
        ref_emb = concept_embedding["ref_embedding_norm"].float()

        # Similarity: [num_templates, num_terms] @ [embed_dim, 1] -> [num_templates, num_terms, 1]
        target_similarity = target_emb @ image_features_norm.T.float()
        ref_similarity = ref_emb @ image_features_norm.T.float()

        # Mean over terms for each template
        target_mean = target_similarity.mean(dim=1).squeeze()  # [num_templates]
        ref_mean = ref_similarity.mean(dim=1).squeeze()  # [num_templates]

        # Softmax between target and reference (contrastive scoring)
        scores = scipy.special.softmax(
            np.array([target_mean.numpy() / temp, ref_mean.numpy() / temp]),
            axis=0
        )

        # Return mean of target scores across templates
        return float(scores[0].mean())

    def extract_features(self, image: Image.Image) -> Dict[str, float]:
        """
        Extract MONET feature scores from a skin lesion image.

        Args:
            image: PIL Image to analyze

        Returns:
            dict with 7 MONET feature scores (0-1 range)
        """
        if not self.loaded:
            self.load()

        # Ensure RGB
        if image.mode != "RGB":
            image = image.convert("RGB")

        # Get image embedding
        image_features = self._encode_image(image)
        image_features_norm = F.normalize(image_features, dim=1)

        # Calculate score for each MONET feature
        features = {}
        for feature_name in MONET_FEATURES:
            concept_emb = self._concept_embeddings[feature_name]
            score = self._calculate_concept_score(image_features_norm, concept_emb)
            features[feature_name] = score

        return features

    def get_feature_vector(self, image: Image.Image) -> List[float]:
        """Get MONET features as a list in the expected order."""
        features = self.extract_features(image)
        return [features[f] for f in MONET_FEATURES]

    def get_feature_tensor(self, image: Image.Image) -> torch.Tensor:
        """Get MONET features as a PyTorch tensor."""
        return torch.tensor(self.get_feature_vector(image), dtype=torch.float32)

    def describe_features(self, features: Dict[str, float], threshold: float = 0.6) -> str:
        """Generate a natural language description of the MONET features."""
        descriptions = {
            "MONET_ulceration_crust": "ulceration or crusting",
            "MONET_hair": "visible hair",
            "MONET_vasculature_vessels": "visible blood vessels",
            "MONET_erythema": "erythema (redness)",
            "MONET_pigmented": "pigmentation",
            "MONET_gel_water_drop_fluid_dermoscopy_liquid": "dermoscopy gel/fluid",
            "MONET_skin_markings_pen_ink_purple_pen": "pen markings",
        }

        present = []
        for feature, score in features.items():
            if score >= threshold:
                desc = descriptions.get(feature, feature)
                present.append(f"{desc} ({score:.0%})")

        if not present:
            # Show top features even if below threshold
            sorted_features = sorted(features.items(), key=lambda x: x[1], reverse=True)[:3]
            present = [f"{descriptions.get(f, f)} ({s:.0%})" for f, s in sorted_features]

        return "Detected features: " + ", ".join(present)

    def analyze(self, image: Image.Image) -> Dict:
        """Full analysis returning features, vector, and description."""
        features = self.extract_features(image)
        vector = [features[f] for f in MONET_FEATURES]
        description = self.describe_features(features)

        return {
            "features": features,
            "vector": vector,
            "description": description,
            "feature_names": MONET_FEATURES,
        }

    def __call__(self, image: Image.Image) -> Dict:
        """Shorthand for analyze()"""
        return self.analyze(image)


# Singleton instance
_monet_instance = None


def get_monet_tool() -> MonetTool:
    """Get or create MONET tool instance"""
    global _monet_instance
    if _monet_instance is None:
        _monet_instance = MonetTool()
    return _monet_instance


if __name__ == "__main__":
    import sys

    print("MONET Tool Test (Correct Implementation)")
    print("=" * 50)

    tool = MonetTool(use_hf=True)
    print("Loading model...")
    tool.load()
    print("Model loaded!")

    if len(sys.argv) > 1:
        image_path = sys.argv[1]
        print(f"\nAnalyzing: {image_path}")
        image = Image.open(image_path).convert("RGB")
        result = tool.analyze(image)

        print("\nMONET Features (Contrastive Scores):")
        for name, score in result["features"].items():
            bar = "█" * int(score * 20)
            print(f"  {name}: {score:.3f} {bar}")

        print(f"\nDescription: {result['description']}")
        print(f"\nVector: {[f'{v:.3f}' for v in result['vector']]}")