File size: 17,349 Bytes
492f6af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
from dataclasses import dataclass
from typing import Optional, Tuple, Dict, Any, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from transformers.utils import ModelOutput
from transformers.modeling_utils import PreTrainedModel
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig, PretrainedConfig
from safetensors.torch import load_file
import torchvision.transforms as transforms
from .build import load_sd_model, load_Florence2_model
from .vlv_utils import initiate_time_steps, normalize, process_caption
from .VLV_stage1 import SDModel, SDConfig
from .configuration_vlv import VLV_Config
import os
import sys
import argparse

def handle_module_prefix(state_dict):
    """Handle 'module.' prefix in state dict keys."""
    if any(k.startswith('module.') for k in state_dict.keys()):
        return {k.replace('module.', ''): v for k, v in state_dict.items()}
    return state_dict

def create_model_args(args):
    """Create model arguments needed by SDModel."""
    model_args = argparse.Namespace()
    model_args.use_text_encoder = args.use_text_encoder
    model_args.batch_size = args.batch_size
    model_args.eval_batch_size = args.batch_size
    model_args.distributed_strategy = 'none'
    model_args.fp32 = args.fp32
    model_args.learnable_token_length = args.learnable_token_length
    model_args.num_inference_steps = args.num_inference_steps
    model_args.image_size = args.image_size
    model_args.guidance_scale = args.guidance_scale
    model_args.unfreeze_florence2_all = False
    model_args.unfreeze_florence2_language_model = False
    model_args.unfreeze_florence2_language_model_decoder = False
    return model_args

def load_model_checkpoint(model, model_path, device):
    """Load model checkpoint."""
    try:
        checkpoint = torch.load(model_path, map_location="cpu")
        
        # Handle different checkpoint formats
        if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint:
            state_dict = checkpoint['model_state_dict']
        elif isinstance(checkpoint, dict) and 'state_dict' in checkpoint:
            state_dict = checkpoint['state_dict']
        else:
            state_dict = checkpoint
            
        state_dict = handle_module_prefix(state_dict)
        missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False)
        
        if missing_keys:
            print(f"Missing keys: {missing_keys[:10]}...")  # Show first 10
        if unexpected_keys:
            print(f"Unexpected keys: {unexpected_keys[:10]}...")  # Show first 10
            
        print(f"Successfully loaded model from {model_path}")
    except Exception as e:
        print(f"Error loading model: {e}")
        raise e

    return model

def initialize_diffusion_model(args):
    """Initialize the diffusion model."""
    config = SDConfig()
    diffusion_model_args = create_model_args(args)
    diffusion_model = SDModel(config, diffusion_model_args)
    _dtype = torch.float32 if diffusion_model_args.fp32 else torch.bfloat16

    # Delete components that aren't needed for inference
    if hasattr(diffusion_model, 'vae'):
        del diffusion_model.vae
    if hasattr(diffusion_model, 'unet'):
        del diffusion_model.unet
    
    # Clear CUDA cache
    torch.cuda.empty_cache()

    diffusion_model = diffusion_model.to(_dtype)
    
    # Freeze parameters that shouldn't be trained
    for param in diffusion_model.language_proj.parameters():
        param.requires_grad = False
    diffusion_model.query_embed.requires_grad = False

    return diffusion_model

class MLP(nn.Module):
    def __init__(self, input_dim, output_dim):
        super(MLP, self).__init__()
        self.layers = nn.Sequential(
            nn.Linear(input_dim, output_dim),
            nn.GELU(),
            nn.Linear(output_dim, output_dim),
        )
        
    def forward(self, x):
        return self.layers(x)


@dataclass
class CLIPDecoderOutput(ModelOutput):
    """
    Output class for the CLIP Decoder model.
    """
    last_hidden_state: Optional[torch.FloatTensor] = None
    generated_ids: Optional[torch.LongTensor] = None
    generated_text: Optional[list] = None


class CLIPDecoder(nn.Module):

    def __init__(
        self, 
        language_model: str,
        VLV_model: SDModel,
        device: torch.device,
        bf16: str,
        qwen2_config: dict = None,
        args: argparse.Namespace = None
    ):
        """
        Initialize the CLIP Decoder model.
        
        Args:
            language_model: Path to the language model
            VLV_model: The VLV model instance
            device: The device to run the model on
            bf16: Whether to use bfloat16 precision
            qwen2_config: Optional qwen2 configuration dict
        """
        super(CLIPDecoder, self).__init__()

        self._dtype = torch.bfloat16 if bf16 == "bf16" else torch.float32
        self.qwen2_tokenizer = AutoTokenizer.from_pretrained(language_model)
        
        self.qwen2_config = AutoConfig.from_pretrained(language_model)
        self.qwen2_model = AutoModelForCausalLM.from_pretrained(
            language_model, 
            torch_dtype=self._dtype, 
            device_map=None, 
            low_cpu_mem_usage=True
        )
        
        self.VLV_model = VLV_model  # fp32 in this case
        self.device = device
        self.mlp = MLP(input_dim=1024, output_dim=self.qwen2_model.config.hidden_size)
        self.ignore_token_id = -100

    
    def get_conditional_context(self, images, batch_size):
        """
        Get conditional context from images using the diffusion model.
        
        Args:
            images: Input images
            batch_size: Batch size
            
        Returns:
            Decoder hidden states from the diffusion model
        """
        prompt = ["<MORE_DETAILED_CAPTION>"] * batch_size
        inputs = self.VLV_model.processor(text=prompt, images=images, return_tensors="pt").to(self.device).to(self._dtype)

        # Ensure all components are on the correct device
        self.VLV_model = self.VLV_model.to(inputs["input_ids"].device)
        self.qwen2_model = self.qwen2_model.to(inputs["input_ids"].device)
        self.mlp = self.mlp.to(inputs["input_ids"].device)
        self.VLV_model.model.language_model.model = self.VLV_model.model.language_model.model.to(inputs["input_ids"].device)
        
        if inputs["input_ids"] is not None:
            inputs_embeds = self.VLV_model.model.language_model.get_input_embeddings()(inputs["input_ids"]).to(self.device)
        
        if inputs["pixel_values"] is not None:
            image_features = self.VLV_model.model._encode_image(inputs["pixel_values"]).to(self.device)
            inputs_embeds, attention_mask = self.VLV_model.model._merge_input_ids_with_image_features(
                image_features, inputs_embeds
            )
        
        if inputs_embeds is not None:
            attention_mask = attention_mask.to(inputs_embeds.dtype)
            
        encoder_outputs = self.VLV_model.model.language_model.model.encoder(
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            output_hidden_states=True,
            return_dict=True
        )
        
        decoder_inputs_embeds = self.VLV_model.query_embed.expand(batch_size, -1, -1)
        decoder_attention_mask = torch.ones(
            (batch_size, self.VLV_model.num_queries),
            dtype=self._dtype,
            device=self.device
        )

        encoder_hidden_states = encoder_outputs.last_hidden_state.to(self._dtype)
        decoder_input_embeds = decoder_inputs_embeds.to(self._dtype)
        attention_mask = attention_mask.to(self._dtype)

        decoder_outputs = self.VLV_model.model.language_model.model.decoder(
            inputs_embeds=decoder_input_embeds,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=attention_mask,
            output_hidden_states=True,
            return_dict=True
        )
            
        return decoder_outputs.last_hidden_state
    
    def process_image(self, images, batch_size):
        """
        Process images to get clip text embeddings.
        
        Args:
            images: Input images
            batch_size: Batch size
            
        Returns:
            Processed clip text embeddings and attention mask
        """
        decoder_hidden_states = self.get_conditional_context(images, batch_size)
        context_embeds = self.VLV_model.language_proj(decoder_hidden_states)
        clip_text_embeds = self.VLV_model.text_encoder(inputs_embeds=context_embeds).last_hidden_state
        clip_text_embeds = self.mlp(clip_text_embeds)
        clip_text_embeds_attention_mask = torch.ones(
            (batch_size, self.VLV_model.num_queries),
            dtype=torch.long,
            device=self.device
        )
        
        return clip_text_embeds, clip_text_embeds_attention_mask
    
    def prepare_generation_inputs(self, clip_text_embeds, clip_text_attention_mask=None):
        """
        Prepare inputs for text generation.
        
        Args:
            clip_text_embeds: Processed clip text embeddings
            clip_text_attention_mask: Attention mask for clip text embeddings
            
        Returns:
            Dictionary of generation inputs
        """
        if clip_text_attention_mask is None:
            clip_text_attention_mask = torch.ones(
                (clip_text_embeds.shape[0], clip_text_embeds.shape[1]),
                dtype=torch.long,
                device=clip_text_embeds.device
            )
            
        return {
            "inputs_embeds": clip_text_embeds,
            "attention_mask": clip_text_attention_mask
        }
    
    def generate(self, images, max_new_tokens=300, num_beams=4, early_stopping=True):
        """
        Generate text from images.
        
        Args:
            images: Input images
            max_new_tokens: Maximum number of tokens to generate
            num_beams: Number of beams for beam search
            early_stopping: Whether to stop early in beam search
            
        Returns:
            CLIPDecoderOutput with generated ids and text
        """
        batch_size = len(images)
        clip_text_embeds, clip_text_attention_mask = self.process_image(images, batch_size)
        generation_inputs = self.prepare_generation_inputs(clip_text_embeds, clip_text_attention_mask)

        generation_inputs["inputs_embeds"] = generation_inputs["inputs_embeds"].to(self._dtype)
        generation_inputs["attention_mask"] = generation_inputs["attention_mask"].to(self._dtype)
    
        generated_ids = self.qwen2_model.generate(
            inputs_embeds=generation_inputs["inputs_embeds"],
            attention_mask=generation_inputs["attention_mask"],
            max_new_tokens=max_new_tokens,
            num_beams=num_beams,
            early_stopping=early_stopping
        )
        
        generated_text = self.qwen2_tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
        processed_generated_text = [process_caption(text) for text in generated_text]
        
        return CLIPDecoderOutput(
            generated_ids=generated_ids,
            generated_text=processed_generated_text
        )
    
    def forward(self, images, captions=None):
        """
        Forward pass for training.
        
        Args:
            images: Input images
            captions: Target captions (optional, for training)
            
        Returns:
            CLIPDecoderOutput with loss and logits
        """
        batch_size = images.shape[0]
        
        # Process images
        clip_text_embeds, clip_text_attention_mask = self.process_image(images, batch_size)
        
        # If no captions provided, return embeddings for generation
        if captions is None:
            return CLIPDecoderOutput(
                last_hidden_state=clip_text_embeds
            )
        
        assert len(captions) == batch_size
        # Process captions for training
        processed_captions = [process_caption(caption) for caption in captions]
        qwen_input_ids = self.qwen2_tokenizer(
            text=processed_captions,
            truncation=True,
            return_tensors="pt",
            padding="max_length",
            max_length=300,
            return_token_type_ids=False,
        ).input_ids
        
        assert len(captions) == batch_size
        qwen_attention_mask = qwen_input_ids.ne(self.qwen2_tokenizer.pad_token_id).to(torch.long).to(self.device)
        
        # Prepare labels for training
        labels = qwen_input_ids
        labels[labels == self.qwen2_tokenizer.pad_token_id] = self.ignore_token_id
        labels = labels.to(self.device)
        
        # Get embeddings for captions to create the full input sequence
        labels_for_embeddings = labels.clone()
        labels_for_embeddings[labels_for_embeddings == self.ignore_token_id] = self.qwen2_tokenizer.pad_token_id
        clip_text_embeds_qwen = self.qwen2_model.get_input_embeddings()(labels_for_embeddings)
        
        # Concatenate the embeddings and prepare attention mask
        inputs_embeds = torch.cat((clip_text_embeds, clip_text_embeds_qwen), dim=1)
        clip_seq_len = clip_text_embeds.shape[1]
        clip_ignore_labels = torch.full((labels.shape[0], clip_seq_len), self.ignore_token_id).to(labels)
        combined_labels = torch.cat((clip_ignore_labels, labels), dim=1)
        
        attention_mask = torch.cat((
            clip_text_attention_mask,
            qwen_attention_mask
        ), dim=1)
        
        # Forward through language model
        outputs = self.qwen2_model(
            inputs_embeds=inputs_embeds,
            labels=combined_labels,
            attention_mask=attention_mask,
            use_cache=False
        )
        return outputs


# HuggingFace Model Wrapper
class VLV_MODEL(PreTrainedModel):
    config_class = VLV_Config
    model_type = "VLV_decoder" 

    def __init__(self, config):
        super().__init__(config)
        """Load the CLIPDecoder model."""
        # Initialize the diffusion model first
        device = "cuda"
        de_diffusion_model = initialize_diffusion_model(config)
        clip_decoder_model = CLIPDecoder(
            language_model=config.qwen_model,
            VLV_model=de_diffusion_model,
            device=device,
            bf16=config.mixed_precision,
            qwen2_config=config.qwen2_config
        )
        
        # Load the trained weights
        # clip_decoder_model = load_model_checkpoint(clip_decoder_model, config.clip_decoder_checkpoint, device)
        
        # Set to evaluation mode
        clip_decoder_model.eval()

        # Store components directly as attributes to match checkpoint structure
        self.VLV_model = clip_decoder_model.VLV_model
        self.qwen2_model = clip_decoder_model.qwen2_model
        self.mlp = clip_decoder_model.mlp
        
        # Keep the full model for methods
        self._clip_decoder_model = clip_decoder_model
        self.max_new_tokens = config.max_length
        self.num_beams = config.num_beams
        self.transform = self.get_transform(config.image_size)

    def get_transform(self, image_size):
        """Transformation pipeline for input images."""
        return transforms.Compose([
            transforms.Resize(image_size),
            transforms.CenterCrop((image_size, image_size)),
            transforms.PILToTensor(),
        ])

    @classmethod
    def from_checkpoint(cls, checkpoint_path, config=None, **kwargs):
        """
        Load model from original training checkpoint.
        
        Args:
            checkpoint_path: Path to the original model.pt checkpoint
            config: Optional VLV_Config, will create default if None
            **kwargs: Additional arguments for model initialization
        """
        if config is None:
            # Create default config
            config = VLV_Config(
                image_size=384,
                guidance_scale=7.5,
                learnable_token_length=77,
                max_length=300,
                num_beams=4,
                **kwargs
            )
        
        # Initialize model
        model = cls(config)
        
        # Load checkpoint weights
        device = "cuda" if torch.cuda.is_available() else "cpu"
        load_model_checkpoint(model._clip_decoder_model, checkpoint_path, device)
        
        return model

    def forward(self, valid_images, max_length):
        valid_images = [self.transform(img) for img in valid_images]
        if hasattr(self._clip_decoder_model, 'module'):
            outputs = self._clip_decoder_model.module.generate(
                valid_images, 
                max_new_tokens=max_length, 
                num_beams=self.num_beams, 
                early_stopping=True
            )
        else:
            outputs = self._clip_decoder_model.generate(
                valid_images, 
                max_new_tokens=max_length, 
                num_beams=self.num_beams, 
                early_stopping=True
            )
        return outputs