File size: 6,359 Bytes
3d02762
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python
"""Experiment 8-v2: stronger unified RGB+pose AfriSign encoder.

This file reuses the Exp8 data pipeline and training loop, but swaps in a
stronger shared encoder:

  - temporal convolutional pose stem before the Transformer
  - language/modality/task-conditioned residual adapters
  - stronger projection head for supervised contrastive learning

It is intentionally a thin wrapper around `exp8_unified_mixed_encoder.py` so the
same manifests, loaders, metrics, and aggregation scripts remain compatible.
"""

from __future__ import annotations

import sys
from pathlib import Path

import torch
import torch.nn as nn

ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))

from experiments import exp8_unified_mixed_encoder as exp8  # noqa: E402


class TemporalPoseStem(nn.Module):
    """TransSLR-style temporal stem for landmark sequences."""

    def __init__(self, feature_dim: int, hidden_dim: int, dropout: float) -> None:
        super().__init__()
        self.in_proj = nn.Linear(feature_dim, hidden_dim)
        self.blocks = nn.ModuleList(
            [
                nn.Sequential(
                    nn.LayerNorm(hidden_dim),
                    nn.Conv1d(hidden_dim, hidden_dim, kernel_size=5, padding=2, groups=hidden_dim),
                    nn.GELU(),
                    nn.Conv1d(hidden_dim, hidden_dim, kernel_size=1),
                    nn.Dropout(dropout),
                )
                for _ in range(3)
            ]
        )
        self.out_norm = nn.LayerNorm(hidden_dim)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.in_proj(x)
        for block in self.blocks:
            y = block[0](x).transpose(1, 2)
            y = block[1](y)
            y = block[2](y)
            y = block[3](y).transpose(1, 2)
            y = block[4](y)
            x = x + y
        return self.out_norm(x)


class ConditionalAdapter(nn.Module):
    """Small residual adapter conditioned by language/modality/level/task context."""

    def __init__(self, hidden_dim: int, bottleneck: int, dropout: float) -> None:
        super().__init__()
        self.norm = nn.LayerNorm(hidden_dim)
        self.down = nn.Linear(hidden_dim, bottleneck)
        self.up = nn.Linear(bottleneck, hidden_dim)
        self.drop = nn.Dropout(dropout)
        self.gate = nn.Sequential(nn.Linear(hidden_dim, hidden_dim), nn.Sigmoid())

    def forward(self, x: torch.Tensor, context: torch.Tensor) -> torch.Tensor:
        update = self.up(torch.nn.functional.gelu(self.down(self.norm(x))))
        gate = self.gate(context).unsqueeze(1)
        return x + self.drop(update * gate)


class StrongUnifiedAfriSignEncoder(nn.Module):
    def __init__(
        self,
        *,
        task_dims: dict[str, int],
        num_languages: int,
        hidden_dim: int,
        feature_dim: int,
        max_tokens: int,
        layers: int,
        heads: int,
        ff_dim: int,
        dropout: float,
        rgb_backbone: str,
        rgb_train_backbone: str,
        rgb_pretrained: bool,
    ) -> None:
        super().__init__()
        self.task_keys = list(task_dims)
        self.pose_stem = TemporalPoseStem(feature_dim, hidden_dim, dropout)
        if rgb_backbone == "small_cnn":
            self.rgb_cnn = exp8.SmallFrameCNN(hidden_dim)
        elif rgb_backbone == "efficientnet_b0":
            self.rgb_cnn = exp8.EfficientNetB0FrameEncoder(hidden_dim, rgb_train_backbone, rgb_pretrained)
        else:
            raise ValueError(f"Unknown rgb_backbone={rgb_backbone!r}")

        self.cls = nn.Parameter(torch.zeros(1, 1, hidden_dim))
        self.pos = nn.Embedding(max_tokens + 1, hidden_dim)
        self.lang_emb = nn.Embedding(num_languages, hidden_dim)
        self.modality_emb = nn.Embedding(len(exp8.MODALITY_IDS), hidden_dim)
        self.level_emb = nn.Embedding(len(exp8.LEVEL_IDS), hidden_dim)
        self.task_emb = nn.Embedding(len(self.task_keys), hidden_dim)

        self.pre_adapter = ConditionalAdapter(hidden_dim, max(32, hidden_dim // 4), dropout)
        enc_layer = nn.TransformerEncoderLayer(
            d_model=hidden_dim,
            nhead=heads,
            dim_feedforward=ff_dim,
            dropout=dropout,
            batch_first=True,
            norm_first=True,
        )
        self.encoder = nn.TransformerEncoder(enc_layer, layers, enable_nested_tensor=False)
        self.post_adapter = ConditionalAdapter(hidden_dim, max(32, hidden_dim // 4), dropout)
        self.norm = nn.LayerNorm(hidden_dim)
        self.drop = nn.Dropout(dropout)
        self.heads = nn.ModuleDict({key: nn.Linear(hidden_dim, dim) for key, dim in task_dims.items()})
        self.projector = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.GELU(),
            nn.LayerNorm(hidden_dim),
            nn.Linear(hidden_dim, hidden_dim),
        )
        nn.init.trunc_normal_(self.cls, std=0.02)

    def encode(self, batch: dict[str, torch.Tensor]) -> torch.Tensor:
        if "pose" in batch:
            x = self.pose_stem(batch["pose"])
        elif "rgb" in batch:
            rgb = batch["rgb"]
            b, t, c, h, w = rgb.shape
            x = self.rgb_cnn(rgb.reshape(b * t, c, h, w)).reshape(b, t, -1)
        else:
            raise ValueError("Batch must contain pose or rgb")

        b, t, _ = x.shape
        positions = torch.arange(t + 1, device=x.device)
        context = (
            self.lang_emb(batch["lang_idx"])
            + self.modality_emb(batch["modality_idx"])
            + self.level_emb(batch["level_idx"])
            + self.task_emb(batch["task_idx"])
        )
        x = torch.cat([self.cls.expand(b, -1, -1), x], dim=1)
        x = x + self.pos(positions).unsqueeze(0) + context.unsqueeze(1)
        x = self.pre_adapter(x, context)
        x = self.encoder(x)
        x = self.post_adapter(x, context)
        return self.drop(self.norm(x[:, 0]))

    def forward(self, batch: dict[str, torch.Tensor], task_key: str) -> torch.Tensor:
        return self.heads[task_key](self.encode(batch))

    def contrast_features(self, features: torch.Tensor) -> torch.Tensor:
        return torch.nn.functional.normalize(self.projector(features), dim=1)


def main() -> None:
    exp8.UnifiedAfriSignEncoder = StrongUnifiedAfriSignEncoder
    exp8.main()


if __name__ == "__main__":
    main()