File size: 13,432 Bytes
264e606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6cbf851
264e606
400535e
264e606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
546c625
264e606
546c625
97233f9
 
 
 
546c625
 
 
 
 
 
264e606
546c625
 
97233f9
546c625
 
 
 
 
97233f9
546c625
9148d56
6a1c9cc
 
97233f9
264e606
 
 
 
 
 
 
 
546c625
97233f9
 
 
 
 
 
 
264e606
 
 
 
 
 
 
546c625
 
97233f9
 
 
 
 
546c625
97233f9
 
546c625
97233f9
264e606
 
97233f9
264e606
 
 
 
 
6a1c9cc
 
264e606
97233f9
 
264e606
 
6a1c9cc
264e606
 
 
 
97233f9
546c625
264e606
 
 
546c625
 
264e606
 
 
 
 
 
c05c346
264e606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch import nn
from torchvision import transforms
from PIL import Image
from transformers import LlamaForCausalLM, LlamaTokenizer, BertModel, BertConfig
from eva_vit import create_eva_vit_g
import requests
from io import BytesIO
import os
from huggingface_hub import hf_hub_download
from transformers import BitsAndBytesConfig
from accelerate import init_empty_weights
import torch
from torch.cuda.amp import autocast
import warnings
MODEL_DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32
token = os.getenv("HF_TOKEN")
import streamlit as st
import torch.nn.functional as F

device = 'cuda' if torch.cuda.is_available() else 'cpu'
class Blip2QFormer(nn.Module):
    def __init__(self, num_query_tokens=32, vision_width=1408):
        super().__init__()
        # Load pre-trained Q-Former config
        self.bert_config = BertConfig(
            vocab_size=30522,
            hidden_size=768,
            num_hidden_layers=12,
            num_attention_heads=12,
            intermediate_size=3072,
            hidden_act="gelu",
            hidden_dropout_prob=0.1,
            attention_probs_dropout_prob=0.1,
            max_position_embeddings=512,
            type_vocab_size=2,
            initializer_range=0.02,
            layer_norm_eps=1e-12,
            pad_token_id=0,
            position_embedding_type="absolute",
            use_cache=True,
            classifier_dropout=None,
        )

        self.bert = BertModel(self.bert_config, add_pooling_layer=False)
        self.query_tokens = nn.Parameter(
            torch.zeros(1, num_query_tokens, self.bert_config.hidden_size)
        )
        self.vision_proj = nn.Linear(vision_width, self.bert_config.hidden_size)

        # Initialize weights
        self._init_weights()

    def _init_weights(self):
        nn.init.normal_(self.query_tokens, std=0.02)
        nn.init.xavier_uniform_(self.vision_proj.weight)
        nn.init.constant_(self.vision_proj.bias, 0)

    def load_from_pretrained(self, url_or_filename):
        if url_or_filename.startswith('http'):
            response = requests.get(url_or_filename)
            checkpoint = torch.load(BytesIO(response.content), map_location='cpu')
        else:
            checkpoint = torch.load(url_or_filename, map_location='cpu')
        state_dict = checkpoint['model'] if 'model' in checkpoint else checkpoint
        msg = self.load_state_dict(state_dict, strict=False)

    def forward(self, visual_features):
        # Project visual features
        with autocast(enabled=False):
            visual_embeds = self.vision_proj(visual_features.float())
        # visual_embeds = self.vision_proj(visual_features.float())
        visual_attention_mask = torch.ones(
            visual_embeds.size()[:-1],
            dtype=torch.long,
            device=visual_embeds.device
        )

        # Expand query tokens
        query_tokens = self.query_tokens.expand(visual_embeds.shape[0], -1, -1)

        # Forward through BERT
        outputs = self.bert(
            input_ids=None,  # No text input
            attention_mask=None,
            inputs_embeds=query_tokens,
            encoder_hidden_states=visual_embeds,
            encoder_attention_mask=visual_attention_mask,
            return_dict=True
        )

        return outputs.last_hidden_state



class SkinGPT4(nn.Module):
    def __init__(self, vit_checkpoint_path,
                 q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth"):
        super().__init__()
        # Image encoder parameters from paper
        self.device = device
        # self.dtype = torch.float16
        self.dtype = MODEL_DTYPE
        self.H, self.W, self.C = 224, 224, 3
        self.P = 14  # Patch size
        self.D = 1408  # ViT embedding dimension
        self.num_query_tokens = 32

        self.vit = self._init_vit(vit_checkpoint_path).to(self.dtype)
        print("Loaded ViT")
        self.ln_vision = nn.LayerNorm(self.D).to(self.dtype)

        self.q_former = Blip2QFormer(
            num_query_tokens=self.num_query_tokens,
            vision_width=self.D
        )
        self.q_former.load_from_pretrained(q_former_model)
        for param in self.q_former.parameters():
            param.requires_grad = False

        print("Loaded QFormer")


        self.llama = self._init_llama()

        self.llama_proj = nn.Linear(
            self.q_former.bert_config.hidden_size,
            self.llama.config.hidden_size
        ).to(self.dtype)

        print(f"Q-Former output dim: {self.q_former.bert_config.hidden_size}")
        print(f"LLaMA input dim: {self.llama.config.hidden_size}")

        for module in [self.vit, self.ln_vision, self.q_former, self.llama_proj, self.llama]:
            for param in module.parameters():
                param.requires_grad = False
            module.eval()

    def _init_vit(self, vit_checkpoint_path):
        """Initialize EVA-ViT-G with paper specifications"""
        vit = create_eva_vit_g(
            img_size=(self.H, self.W),
            patch_size=self.P,
            embed_dim=self.D,
            depth=39,
            num_heads=16,
            mlp_ratio=4.3637,
            qkv_bias=True,
            drop_path_rate=0.1,
            norm_layer=nn.LayerNorm,
            init_values=1e-5
        ).to(self.dtype)
        if not hasattr(vit, 'norm'):
            vit.norm = nn.LayerNorm(self.D)
        checkpoint = torch.load(vit_checkpoint_path, map_location='cpu')
        # 3. Filter weights for ViT components only
        vit_weights = {k.replace("vit.", ""): v
                       for k, v in checkpoint.items()
                       if k.startswith("vit.")}

        # 4. Load weights while ignoring classifier head
        vit.load_state_dict(vit_weights, strict=False)


        return vit.eval()

    def _init_llama(self):
        """Initialize frozen LLaMA-2-13b-chat with proper error handling"""
        try:
            device_map = {
                "": 0 if torch.cuda.is_available() else "cpu"
            }
            # First try loading with device_map="auto"
            model = LlamaForCausalLM.from_pretrained(
                "meta-llama/Llama-2-13b-chat-hf",
                token=token,
                torch_dtype=torch.float16,
                device_map=device_map,
                low_cpu_mem_usage=True
            )

            return model.eval()

        except Exception as e:
            raise ImportError(
                f"Failed to load LLaMA model. Please ensure:\n"
                f"1. You have accepted the license at: https://huggingface.co/meta-llama/Llama-2-13b-chat-hf\n"
                f"2. Your Hugging Face token is correct\n"
                f"3. Required packages are installed: pip install accelerate bitsandbytes transformers\n"
                f"Original error: {str(e)}"
            )

    def encode_image(self, x):
        """Convert image to patch embeddings following Eq. (1)"""
        # x: (B, C, H, W)
        x = x.to(self.dtype)
        if x.dim() == 3:
            x = x.unsqueeze(0)  # Add batch dimension if missing
        if x.dim() != 4:
            raise ValueError(f"Input must be 4D tensor (got {x.dim()}D)")

        B, C, H, W = x.shape
        N = (H * W) // (self.P ** 2)

        x = self.vit.patch_embed(x)

        num_patches = x.shape[1]
        pos_embed = self.vit.pos_embed[:, 1:num_patches + 1, :]
        x = x + pos_embed

        # Add class token
        class_token = self.vit.cls_token.expand(x.shape[0], -1, -1)
        x = torch.cat([class_token, x], dim=1)
        for blk in self.vit.blocks:
            x = blk(x)
        x = self.vit.norm(x)
        vit_features = self.ln_vision(x)

        # Q-Former forward pass
        with torch.no_grad():
            qformer_output = self.q_former(vit_features.float())
            image_embeds = self.llama_proj(qformer_output.to(self.dtype))

        return image_embeds

    def generate(self, images, user_input=None, max_new_tokens=300):

        image_embeds = self.encode_image(images)

        print(f"Aligned features : {image_embeds}")
        print(f"\n Images embeddings shape : {image_embeds.shape} \n Llama config hidden size : {self.llama.config.hidden_size}")

        print(
            f"\n[VALIDATION] Visual embeds - Mean: {image_embeds.mean().item():.4f}, Std: {image_embeds.std().item():.4f}")

        if image_embeds.shape[-1] != self.llama.config.hidden_size:
            raise ValueError(
                f"Feature dimension mismatch. "
                f"Q-Former output: {image_embeds.shape[-1]}, "
                f"LLaMA expected: {self.llama.config.hidden_size}"
            )


        # prompt = (
        #     "### Instruction: <Img><IMAGE></Img> "
        #     "Could you describe the skin condition in this image? "
        #     "Please provide a detailed analysis including possible diagnoses. "
        #     "### Response:"
        # )

        prompt = """### Skin Diagnosis Analysis ###
        <IMAGE>
        Could you describe the skin condition in this image?
        Please provide a detailed analysis including possible diagnoses.
        ### Response:"""


        print(f"\n[DEBUG] Raw Prompt:\n{prompt}")

        self.tokenizer = LlamaTokenizer.from_pretrained(
            "meta-llama/Llama-2-13b-chat-hf",
            token=token,
            padding_side="right"
        )
        # self.tokenizer.add_special_tokens({'additional_special_tokens': ['<Img>', '</Img>', '<ImageHere>']})
        num_added = self.tokenizer.add_special_tokens({
            'additional_special_tokens': ['<IMAGE>']
        })

        if num_added == 0:
            raise ValueError("Failed to add <IMAGE> token!")

        self.llama.resize_token_embeddings(len(self.tokenizer))

        inputs = self.tokenizer(prompt, return_tensors="pt").to(images.device)

        print(f"\n[DEBUG] Tokenized input IDs:\n{inputs.input_ids}")
        print(f"[DEBUG] Special token positions: {self.tokenizer.all_special_tokens}")

        # Prepare embeddings
        input_embeddings = self.llama.model.embed_tokens(inputs.input_ids)
        visual_embeds = image_embeds.mean(dim=1)

        # image_token_id = self.tokenizer.convert_tokens_to_ids("<ImageHere>")
        image_token_id = self.tokenizer.convert_tokens_to_ids("<IMAGE>")
        replace_positions = (inputs.input_ids == image_token_id).nonzero()

        if len(replace_positions) == 0:
            raise ValueError("No <IMAGE> tokens found in prompt!")

        if len(replace_positions[0]) == 0:
            raise ValueError("Image token not found in prompt")

        print(f"\n[DEBUG] Image token found at position: {replace_positions}")


        print(f"\n[DEBUG] Before replacement:")
        print(f"Text embeddings shape: {input_embeddings.shape}")
        print(f"Visual embeddings shape: {visual_embeds.shape}")
        print(f"Image token at {replace_positions[0][1].item()}:")
        print(f"Image token embedding (before):\n{input_embeddings[0, replace_positions[0][1], :5]}...")

        for pos in replace_positions:
            input_embeddings[0, pos[1]] = visual_embeds[0]

        print(f"\n[DEBUG] After replacement:")
        print(f"Image token embedding (after):\n{input_embeddings[0, replace_positions[0][1], :5]}...")

        outputs = self.llama.generate(
            inputs_embeds=input_embeddings,
            max_new_tokens=max_new_tokens,
            temperature=0.7,
            top_k=40,
            top_p=0.9,
            repetition_penalty=1.1,
            do_sample=True,
            pad_token_id = self.tokenizer.eos_token_id,
            eos_token_id = self.tokenizer.eos_token_id
        )


        full_output = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
        print(f"Full Output from llama : {full_output}")
        response = full_output.split("### Response:")[-1].strip()
        # print(f"Response from llama : {full_output}")

        return response


class SkinGPTClassifier:
    def __init__(self, device='cuda' if torch.cuda.is_available() else 'cpu'):
        self.device = torch.device(device)
        self.conversation_history = []

        with st.spinner("Loading AI models (this may take several minutes)..."):
            self.model = self._load_model()
        # print(f"Q-Former output shape: {self.model.q_former(torch.randn(1, 197, 1408)).shape}")
        # print(f"Projection layer: {self.model.llama_proj}")

        # Image transformations
        self.transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])

    def _load_model(self):
        model_path = hf_hub_download(
            repo_id="KeerthiVM/SkinCancerDiagnosis",
            filename="dermnet_finetuned_version1.pth",
        )
        model = SkinGPT4(vit_checkpoint_path=model_path).eval()
        model = model.to(self.device)
        return model

    def predict(self, image):
        image = image.convert('RGB')
        image_tensor = self.transform(image).unsqueeze(0).to(self.device)
        with torch.no_grad():
            diagnosis = self.model.generate(image_tensor)

        return {
            "diagnosis": diagnosis,
            "visual_features": None  # Can return features if needed
        }