File size: 8,724 Bytes
0adab2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""A2C2 correction head architecture for cached BEHAVIOR/OpenPI features."""

from __future__ import annotations

from dataclasses import dataclass
import math

import torch
from torch import Tensor, nn


@dataclass(frozen=True)
class A2C2CorrectionHeadConfig:
    state_dim: int = 256
    action_dim: int = 23
    action_horizon: int = 32
    base_policy_z_dim: int = 2048
    use_base_policy_z: bool = True
    time_dim: int = 2

    dim_model: int = 512
    n_heads: int = 8
    n_encoder_layers: int = 6
    dim_feedforward: int = 2048
    dropout: float = 0.1
    mlp_hidden_dim: int = 1024


def _sinusoidal_positions(length: int, dim: int) -> Tensor:
    if dim % 2 != 0:
        raise ValueError("dim must be even for sinusoidal positional encoding.")

    position = torch.arange(length, dtype=torch.float32).unsqueeze(1)
    div_term = torch.exp(torch.arange(0, dim, 2, dtype=torch.float32) * (-math.log(10000.0) / dim))
    pe = torch.zeros(length, dim, dtype=torch.float32)
    pe[:, 0::2] = torch.sin(position * div_term)
    pe[:, 1::2] = torch.cos(position * div_term)
    return pe


class A2C2CorrectionHead(nn.Module):
    """Transformer + MLP correction head following the A2C2 residual design."""

    def __init__(self, config: A2C2CorrectionHeadConfig | None = None) -> None:
        super().__init__()
        self.config = config or A2C2CorrectionHeadConfig()
        cfg = self.config

        self.cls_token = nn.Parameter(torch.zeros(1, 1, cfg.dim_model))
        self.type_embedding = nn.Parameter(torch.zeros(6, cfg.dim_model))

        self.state_proj = nn.Linear(cfg.state_dim, cfg.dim_model)
        if cfg.use_base_policy_z:
            self.z_proj = nn.Linear(cfg.base_policy_z_dim, cfg.dim_model)
        self.time_proj = nn.Linear(cfg.time_dim, cfg.dim_model)
        self.action_proj = nn.Linear(cfg.action_dim, cfg.dim_model)

        chunk_pos = _sinusoidal_positions(cfg.action_horizon, cfg.dim_model)
        self.register_buffer("chunk_pos_embedding", chunk_pos, persistent=False)

        encoder_layer = nn.TransformerEncoderLayer(
            d_model=cfg.dim_model,
            nhead=cfg.n_heads,
            dim_feedforward=cfg.dim_feedforward,
            dropout=cfg.dropout,
            activation="gelu",
            batch_first=True,
            norm_first=True,
        )
        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=cfg.n_encoder_layers)
        self.encoder_norm = nn.LayerNorm(cfg.dim_model)

        head_input_token_count = 5 if cfg.use_base_policy_z else 4
        head_input_dim = cfg.dim_model * head_input_token_count + cfg.action_dim
        self.residual_head = nn.Sequential(
            nn.Linear(head_input_dim, cfg.mlp_hidden_dim),
            nn.GELU(),
            nn.Dropout(cfg.dropout),
            nn.Linear(cfg.mlp_hidden_dim, cfg.mlp_hidden_dim),
            nn.GELU(),
            nn.Dropout(cfg.dropout),
            nn.Linear(cfg.mlp_hidden_dim, cfg.action_dim),
        )

        self._reset_parameters()

    @staticmethod
    def make_time_feature(chunk_index: Tensor, horizon: int) -> Tensor:
        """Create [sin, cos] phase features from chunk indices."""

        idx = chunk_index.to(dtype=torch.float32)
        denom = max(horizon - 1, 1)
        phase = 2.0 * math.pi * idx / denom
        return torch.stack([torch.sin(phase), torch.cos(phase)], dim=-1)

    def forward(
        self,
        observation_state: Tensor,
        selected_base_action: Tensor,
        base_action_chunk: Tensor,
        base_policy_z: Tensor,
        time_feature: Tensor,
        valid_action_mask: Tensor | None = None,
    ) -> Tensor:
        """Predict residual action delta.

        Args:
            observation_state: [B, state_dim]
            selected_base_action: [B, action_dim], the current action being corrected.
            base_action_chunk: [B, H, action_dim]
            base_policy_z: [B, z_dim]
            time_feature: [B, 2]
            valid_action_mask: optional bool tensor [B, H], True for valid chunk
                entries. Invalid chunk entries are ignored by transformer attention.

        Returns:
            Tensor [B, action_dim], the predicted residual delta.
        """

        cfg = self.config
        batch_size = observation_state.shape[0]
        device = observation_state.device
        dtype = observation_state.dtype

        self._validate_inputs(observation_state, selected_base_action, base_action_chunk, base_policy_z, time_feature)

        cls = self.cls_token.to(device=device, dtype=dtype).expand(batch_size, -1, -1)
        cls = cls + self.type_embedding[0].to(device=device, dtype=dtype)

        state_token = self.state_proj(observation_state).unsqueeze(1)
        state_token = state_token + self.type_embedding[1].to(device=device, dtype=dtype)

        time_token = self.time_proj(time_feature).unsqueeze(1)
        time_token = time_token + self.type_embedding[3].to(device=device, dtype=dtype)

        selected_action_token = self.action_proj(selected_base_action).unsqueeze(1)
        selected_action_token = selected_action_token + self.type_embedding[4].to(device=device, dtype=dtype)

        chunk_tokens = self.action_proj(base_action_chunk)
        chunk_pos = self.chunk_pos_embedding[: base_action_chunk.shape[1]].to(device=device, dtype=dtype)
        chunk_tokens = chunk_tokens + chunk_pos.unsqueeze(0)
        chunk_tokens = chunk_tokens + self.type_embedding[5].to(device=device, dtype=dtype)

        prefix_tokens = [cls, state_token]
        if cfg.use_base_policy_z:
            z_token = self.z_proj(base_policy_z).unsqueeze(1)
            z_token = z_token + self.type_embedding[2].to(device=device, dtype=dtype)
            prefix_tokens.append(z_token)
        prefix_tokens.extend([time_token, selected_action_token])

        tokens = torch.cat([*prefix_tokens, chunk_tokens], dim=1)

        padding_mask = None
        if valid_action_mask is not None:
            valid_action_mask = valid_action_mask.to(device=device, dtype=torch.bool)
            prefix_mask = torch.zeros(batch_size, len(prefix_tokens), device=device, dtype=torch.bool)
            padding_mask = torch.cat([prefix_mask, ~valid_action_mask], dim=1)

        encoded = self.encoder(tokens, src_key_padding_mask=padding_mask)
        encoded = self.encoder_norm(encoded)

        cls_state = encoded[:, 0]
        state_state = encoded[:, 1]
        if cfg.use_base_policy_z:
            z_state = encoded[:, 2]
            time_state = encoded[:, 3]
            selected_action_state = encoded[:, 4]
            head_states = [cls_state, state_state, z_state, time_state, selected_action_state]
        else:
            time_state = encoded[:, 2]
            selected_action_state = encoded[:, 3]
            head_states = [cls_state, state_state, time_state, selected_action_state]

        head_input = torch.cat(
            [*head_states, selected_base_action],
            dim=-1,
        )
        return self.residual_head(head_input)

    def _validate_inputs(
        self,
        observation_state: Tensor,
        selected_base_action: Tensor,
        base_action_chunk: Tensor,
        base_policy_z: Tensor,
        time_feature: Tensor,
    ) -> None:
        cfg = self.config
        if observation_state.ndim != 2 or observation_state.shape[-1] != cfg.state_dim:
            raise ValueError(f"observation_state must have shape [B, {cfg.state_dim}].")
        if selected_base_action.ndim != 2 or selected_base_action.shape[-1] != cfg.action_dim:
            raise ValueError(f"selected_base_action must have shape [B, {cfg.action_dim}].")
        if base_action_chunk.ndim != 3 or base_action_chunk.shape[-1] != cfg.action_dim:
            raise ValueError(f"base_action_chunk must have shape [B, H, {cfg.action_dim}].")
        if base_action_chunk.shape[1] > cfg.action_horizon:
            raise ValueError(f"base_action_chunk horizon cannot exceed {cfg.action_horizon}.")
        if cfg.use_base_policy_z and (base_policy_z.ndim != 2 or base_policy_z.shape[-1] != cfg.base_policy_z_dim):
            raise ValueError(f"base_policy_z must have shape [B, {cfg.base_policy_z_dim}].")
        if time_feature.ndim != 2 or time_feature.shape[-1] != cfg.time_dim:
            raise ValueError(f"time_feature must have shape [B, {cfg.time_dim}].")

    def _reset_parameters(self) -> None:
        nn.init.trunc_normal_(self.cls_token, std=0.02)
        nn.init.trunc_normal_(self.type_embedding, std=0.02)
        for module in self.modules():
            if isinstance(module, nn.Linear):
                nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.zeros_(module.bias)