File size: 9,710 Bytes
8d7b4b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import torch
import torch.nn as nn
from typing import Optional

from src.models.encoders import (
    AttentionPooling,
    BiGRUEncoder,
    ConvGRUEncoder,
    TemporalConvEncoder,
    TransformerEncoder,
)
from src.models.fusion import MultimodalFusion
from src.models.heads import ClassificationHead, ConfidenceHead, QuestionHead, RegressionHead


class MultimodalDepressionModel(nn.Module):
    def __init__(
        self,
        audio_dim: int = 512,
        video_dim: int = 512,
        text_dim: int = 384,
        fusion_dim: int = 256,
        num_questions: int = 8,
        encoder_type: str = "transformer",
        pooling_type: str = "attention",
        num_heads: int = 4,
        num_layers: int = 2,
        dropout: float = 0.2,
        fusion_mode: str = "hybrid",
        modality_dropout: float = 0.1,
    ):
        super().__init__()
        self.encoder_type = encoder_type
        self.pooling_type = pooling_type
        self.audio_input_dim = audio_dim
        self.video_input_dim = video_dim
        self.text_input_dim = text_dim

        self.audio_encoder = self._build_encoder(audio_dim, fusion_dim, dropout, num_layers, num_heads)
        self.video_encoder = self._build_encoder(video_dim, fusion_dim, dropout, num_layers, num_heads)
        self.text_encoder = self._build_encoder(text_dim, fusion_dim, dropout, num_layers, num_heads)

        self.audio_projection = self._build_projection(audio_dim, fusion_dim)
        self.video_projection = self._build_projection(video_dim, fusion_dim)
        self.text_projection = self._build_projection(text_dim, fusion_dim)

        self.audio_missing = nn.Parameter(torch.zeros(fusion_dim))
        self.video_missing = nn.Parameter(torch.zeros(fusion_dim))
        self.text_missing = nn.Parameter(torch.zeros(fusion_dim))

        # Persistent attention poolers — one per modality so weights ARE trained.
        # (previously created fresh each forward pass → random, never trained → CCC gap)
        if self.pooling_type == "attention":
            self.audio_pooler = AttentionPooling(fusion_dim)
            self.video_pooler = AttentionPooling(fusion_dim)
            self.text_pooler  = AttentionPooling(fusion_dim)

        self.fusion = MultimodalFusion(
            embed_dim=fusion_dim,
            fusion_dim=fusion_dim,
            mode=fusion_mode,
            dropout=dropout,
        )
        self.phq_head = RegressionHead(fusion_dim)
        self.question_head = QuestionHead(fusion_dim, num_questions)
        self.classification_head = ClassificationHead(fusion_dim)
        self.confidence_head = ConfidenceHead(fusion_dim)

        self.modality_dropout = modality_dropout

    def _build_encoder(
        self,
        input_dim: int,
        hidden_dim: int,
        dropout: float,
        num_layers: int,
        num_heads: int,
    ) -> nn.Module:
        if self.encoder_type == "conv":
            return TemporalConvEncoder(input_dim, hidden_dim, dropout=dropout)
        if self.encoder_type == "conv_gru":
            return ConvGRUEncoder(input_dim, hidden_dim, num_layers=num_layers, dropout=dropout)
        if self.encoder_type == "bigru":
            return BiGRUEncoder(input_dim, hidden_dim, num_layers=num_layers, dropout=dropout)
        return TransformerEncoder(input_dim, num_layers=num_layers, num_heads=num_heads, dropout=dropout)

    def _build_projection(self, input_dim: int, hidden_dim: int) -> nn.Module:
        if self.encoder_type == "bigru":
            return nn.Linear(hidden_dim * 2, hidden_dim)
        if self.encoder_type == "transformer":
            return nn.Linear(input_dim, hidden_dim)
        return nn.Linear(hidden_dim, hidden_dim)

    @staticmethod
    def _mean_pool(x: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
        mask = mask.unsqueeze(-1).float()
        summed = (x * mask).sum(dim=1)
        length = mask.sum(dim=1).clamp(min=1.0)
        return summed / length

    def _pooled_representation(
        self, x: torch.Tensor, mask: torch.Tensor, pooler: Optional[nn.Module] = None
    ) -> torch.Tensor:
        if self.pooling_type == "attention" and pooler is not None:
            return pooler(x, mask)
        return self._mean_pool(x, mask)

    def _encode_modality(
        self,
        x: Optional[torch.Tensor],
        mask: Optional[torch.Tensor],
        encoder: nn.Module,
        projector: nn.Module,
        missing_token: nn.Parameter,
        pooler: Optional[nn.Module] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if x is None or x.numel() == 0 or mask is None:
            batch = x.shape[0] if x is not None else 1
            return missing_token.unsqueeze(0).expand(batch, -1), torch.zeros(batch, dtype=torch.bool, device=missing_token.device)

        mask = mask.to(dtype=torch.bool)
        present = mask.any(dim=1)
        encoded = encoder(x, mask)
        pooled = self._pooled_representation(encoded, mask, pooler)
        projected = projector(pooled)

        if not present.all():
            fallback = missing_token.unsqueeze(0).expand(projected.size(0), -1)
            projected = projected * present.unsqueeze(-1).to(projected.dtype) + fallback * (~present).unsqueeze(-1).to(projected.dtype)

        return projected, present

    def _drop_modality(
        self,
        modality_repr: torch.Tensor,
        present: torch.Tensor,
        missing_token: nn.Parameter,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if not self.training or self.modality_dropout <= 0.0:
            return modality_repr, present

        keep = torch.rand(modality_repr.size(0), device=modality_repr.device) > self.modality_dropout
        keep = keep & present
        if keep.all():
            return modality_repr, present

        fallback = missing_token.unsqueeze(0).expand_as(modality_repr)
        modality_repr = modality_repr * keep.unsqueeze(-1).to(modality_repr.dtype) + fallback * (~keep).unsqueeze(-1).to(modality_repr.dtype)
        return modality_repr, keep

    def forward(
        self,
        audio: Optional[torch.Tensor],
        video: Optional[torch.Tensor],
        text: Optional[torch.Tensor],
        audio_mask: Optional[torch.Tensor],
        video_mask: Optional[torch.Tensor],
        text_mask: Optional[torch.Tensor],
    ) -> dict[str, torch.Tensor]:
        audio_repr, audio_present = self._encode_modality(
            audio,
            audio_mask,
            self.audio_encoder,
            self.audio_projection,
            self.audio_missing,
            pooler=getattr(self, "audio_pooler", None),
        )
        video_repr, video_present = self._encode_modality(
            video,
            video_mask,
            self.video_encoder,
            self.video_projection,
            self.video_missing,
            pooler=getattr(self, "video_pooler", None),
        )
        text_repr, text_present = self._encode_modality(
            text,
            text_mask,
            self.text_encoder,
            self.text_projection,
            self.text_missing,
            pooler=getattr(self, "text_pooler", None),
        )

        if self.training and self.modality_dropout > 0.0:
            a_keep = torch.rand(audio_repr.size(0), device=audio_repr.device) > self.modality_dropout
            v_keep = torch.rand(video_repr.size(0), device=video_repr.device) > self.modality_dropout
            t_keep = torch.rand(text_repr.size(0), device=text_repr.device) > self.modality_dropout
            
            # If all are false, force one to be true among the ones that are actually present
            a_keep = a_keep & audio_present
            v_keep = v_keep & video_present
            t_keep = t_keep & text_present
            all_dropped = ~(a_keep | v_keep | t_keep)
            
            # For rows where all dropped, try text, then audio, then video
            if all_dropped.any():
                t_keep = t_keep | (all_dropped & text_present)
                all_dropped_still = all_dropped & ~text_present
                a_keep = a_keep | (all_dropped_still & audio_present)
                all_dropped_still = all_dropped_still & ~audio_present
                v_keep = v_keep | (all_dropped_still & video_present)
            
            def apply_drop(repr_t, present_t, keep_t, missing_t):
                if keep_t.all(): return repr_t, present_t
                fallback = missing_t.unsqueeze(0).expand_as(repr_t)
                repr_out = repr_t * keep_t.unsqueeze(-1).to(repr_t.dtype) + fallback * (~keep_t).unsqueeze(-1).to(repr_t.dtype)
                return repr_out, keep_t
                
            audio_repr, audio_present = apply_drop(audio_repr, audio_present, a_keep, self.audio_missing)
            video_repr, video_present = apply_drop(video_repr, video_present, v_keep, self.video_missing)
            text_repr, text_present = apply_drop(text_repr, text_present, t_keep, self.text_missing)

        modality_present = torch.stack([audio_present, video_present, text_present], dim=1)
        fused, modality_scores, entropy, confidence = self.fusion(
            audio_repr,
            video_repr,
            text_repr,
            modality_present=modality_present,
        )

        return {
            "phq_total": self.phq_head(fused).squeeze(-1),
            "phq_questions": self.question_head(fused),
            "classification": self.classification_head(fused).squeeze(-1),
            "confidence": self.confidence_head(fused).squeeze(-1),
            "modality_scores": modality_scores,
            "modality_confidence": confidence,
            "fusion_entropy": entropy,
            "entropy": entropy,
        }