File size: 9,935 Bytes
b6faa73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b0ea89d
b6faa73
 
 
ce081c7
8948365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b6faa73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import warnings
from transformers import (
    HubertModel, 
    AutoProcessor, 
    AutoTokenizer, 
    AutoModel
)
warnings.filterwarnings("ignore")
import torchvision.transforms as transforms
from PIL import Image
#################################################################
#                   Audio Embedder
#################################################################
class AudioEmbedder(nn.Module):
    """
    Pre-trained HuBERT (or similar) to extract audio features from raw audio (16kHz).
    Projects them down to a desired embedding dimension.
    """
    def __init__(self, embedding_dim=512, hubert_name="facebook/hubert-base-ls960"):
        super().__init__()
        self.processor = AutoProcessor.from_pretrained("facebook/hubert-large-ls960-ft")  
        self.hubert = HubertModel.from_pretrained(hubert_name)
        self.projection = nn.Linear(self.hubert.config.hidden_size, embedding_dim)
        
        for param in self.hubert.parameters():
            param.requires_grad = True
        for param in self.projection.parameters():
            param.requires_grad = True
        
    def forward(self, audio_input: torch.Tensor) -> torch.Tensor:
        """
        Args:
            audio_input: (B, T) raw audio waveform at 16kHz
            
        Returns:
            audio_feats: (B, Na, D) 
                B = batch size
                Na = number of audio tokens (T/320 for Hubert)
                D = embedding_dim
        """
        if len(audio_input.shape) == 3:  # shape: [B, 1, T]
            audio_input = audio_input.squeeze(0)  # squeeze first dim to get [B, T]
        inputs = self.processor(
            audio_input, 
            return_tensors="pt",
            sampling_rate=16000,
            padding=True,
            return_attention_mask=True
        ).input_values.squeeze(0)
        device = next(self.parameters()).device
        inputs = inputs.to(device)
        
        hubert_output = self.hubert(inputs).last_hidden_state  # (B, T', hidden_size)
        
        audio_feats = self.projection(hubert_output)  # (B, T', D)
        
        return audio_feats


#################################################################
#                   Text Embedder
#################################################################
class TextEmbedder(nn.Module):
    """
    Pre-trained BERT-like model (ModernBERT or similar) to extract text features.
    Projects them down to a desired embedding dimension.
    """
    def __init__(self, embedding_dim=512, model_name="answerdotai/ModernBERT-base"):
        super().__init__()
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.encoder = AutoModel.from_pretrained(model_name)
        self.projection = nn.Linear(self.encoder.config.hidden_size, embedding_dim)
        print("Using text model: ", model_name)
        
        for param in self.encoder.parameters():
            param.requires_grad = True
        for param in self.projection.parameters():
            param.requires_grad = True
        
    def forward(self, text_list):
        """
        Args:
            text_list: List[str], batch of text inputs
            
        Returns:
            text_feats: (B, Nt, D)
            attention_mask: (B, Nt)
        """
        inputs = self.tokenizer(
            text_list, 
            padding=True,
            truncation=True,
            add_special_tokens=False,
            max_length=128,
            return_tensors="pt"
        )
        device = next(self.parameters()).device
        for k in inputs:
            inputs[k] = inputs[k].to(device)

        outputs = self.encoder(**inputs)  # (B, Nt, hidden_size)
        hidden_states = outputs.last_hidden_state
        text_feats = self.projection(hidden_states)  # (B, Nt, D)
        
        return text_feats, inputs["attention_mask"]


#################################################################
#                   Visual Embedder
#################################################################
class ViTEmbedder(nn.Module):
    """
    DINOv2 to extract patch embeddings from an image.
    Then projects to a common dimension with a linear layer.
    """
    def __init__(self, model_name='facebookresearch/dinov2', arch='dinov2_vitb14',
                 embedding_dim=512, dropout_prob=0.1):
        super().__init__()
        self.model = torch.hub.load(model_name, arch)
        print("Using DINOv2 model: ", arch)
        self.projection = nn.Linear(self.model.embed_dim, embedding_dim)
        self.dropout = nn.Dropout(p=dropout_prob)

        for param in self.model.parameters():
            param.requires_grad = True

    def forward(self, x):
        """
        Args:
            x: (B, 3, H, W), e.g. (B,3,224,224) image batch
        Returns:
            visual_feats: (B, Nv, D)
                Nv = number of visual tokens
                D  = embedding_dim
        """
        if len(x.shape) == 5:  # shape: [1, 1, 3, 224, 224]
            x = x.squeeze(0)  # get [1, 3, 224, 224]
        if len(x.shape) == 3:
            x = x.unsqueeze(0)
        patches = self.model.get_intermediate_layers(x, n=1)[0]  
        feats = self.projection(patches)
        feats = self.dropout(feats)
        
        return feats

class Triad(nn.Module):
    def __init__(
        self, 
        audio_model_name="facebook/hubert-base-ls960",
        text_model_name="distilbert/distilbert-base-uncased",
        temperature=2.0,
        patch_sparsity_threshold=0.3,
        patch_sparsity_weight=0.1,
        visual_dropout_prob=0.1
    ):
        super().__init__()

        self.audio_embedder = AudioEmbedder(embedding_dim=512, hubert_name=audio_model_name)
        self.text_embedder  = TextEmbedder(embedding_dim=512, model_name=text_model_name)
        self.visual_embedder = ViTEmbedder(arch='dinov2_vitb14',
                                           embedding_dim=512,
                                           dropout_prob=visual_dropout_prob)

        self.temperature = nn.Parameter(torch.tensor(temperature))
        self.patch_sparsity_threshold = patch_sparsity_threshold
        self.patch_sparsity_weight = patch_sparsity_weight

    def compute_similarity_matrix(self, feats1, feats2):
        """
        Generic token-level dot-product similarity between feats1 and feats2.
        feats1: (B, N1, D)
        feats2: (B, N2, D)
        Returns sim: (B, N1, N2)
        """
        sim = torch.bmm(feats1, feats2.transpose(1, 2))
        return sim / self.temperature
        
    def forward(self, image=None, audio=None, text_list=None):
        assert image is not None or audio is not None or text_list is not None, "At least one modality must be provided"
        if image is not None: assert image is not str, "Frames should be a path to an image"
        if audio is not None: 
            assert isinstance(audio, torch.Tensor) and len(audio.shape) == 2, "Audio must be a PyTorch tensor of shape (B, T)"
        if text_list is not None:
            assert isinstance(text_list, list) and len(text_list) == 1, "Text list must be a list of strings of length 1"
        if image is not None:
            device = next(self.parameters()).device
            
            # Handle batch of file paths
            if isinstance(image, list):
                # Process a list of image paths
                processed_images = []
                for img_path in image:
                    img = Image.open(img_path).convert('RGB')
                    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])
                    ])
                    processed_img = transform(img).to(device)
                    processed_images.append(processed_img)
                image = torch.stack(processed_images, dim=0)  # [B, 3, 224, 224]
            
            # Handle single file path
            elif isinstance(image, str):
                img = Image.open(image).convert('RGB')
                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])
                ])
                image = transform(img).to(device).unsqueeze(0)  # Add batch dimension [1, 3, 224, 224]
            
            # Handle tensor input (assume it's already processed but may need device transfer)
            elif isinstance(image, torch.Tensor):
                # If single image without batch dimension
                if image.dim() == 3:
                    image = image.unsqueeze(0)  # Add batch dimension
                image = image.to(device)
                
        embeddings = {}
        if image is not None:
            embeddings['visual_feats'] = self.visual_embedder(image)
        if audio is not None:
            embeddings['audio_feats'] = self.audio_embedder(audio)
        if text_list is not None:
            embeddings['text_feats'], _ = self.text_embedder(text_list)
        # if two or more modalities are present, we compute the similarity matrix 
        if image is not None and text_list is not None:
            embeddings['vis_text_sim_matrix'] = self.compute_similarity_matrix(embeddings['text_feats'], embeddings['visual_feats'])
        if audio is not None and image is not None:
            embeddings['vis_audio_sim_matrix'] = self.compute_similarity_matrix(embeddings['audio_feats'], embeddings['visual_feats'])
        if text_list is not None and audio is not None:
            embeddings['text_audio_sim_matrix'] = self.compute_similarity_matrix(embeddings['text_feats'], embeddings['audio_feats'])
        return embeddings