MCPcopy Create free account
hub / github.com/SJTU-DENG-Lab/WLA / ActionEncoder

Class ActionEncoder

models/action_model/action_encoder.py:57–101  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

55
56
57class ActionEncoder(nn.Module):
58 def __init__(self, action_dim, hidden_size):
59 super().__init__()
60 self.hidden_size = hidden_size
61
62 # W1: R^{w x d}, W2: R^{w x 2w}, W3: R^{w x w}
63 self.W1 = nn.Linear(action_dim, hidden_size) # (d -> w)
64 self.W2 = nn.Linear(2 * hidden_size, hidden_size) # (2w -> w)
65 self.W3 = nn.Linear(hidden_size, hidden_size) # (w -> w)
66
67 self.pos_encoding = SinusoidalPositionalEncoding(hidden_size)
68
69 def forward(self, actions, timesteps):
70 """
71 actions: shape (B, T, action_dim)
72 timesteps: shape (B,) -- a single scalar per batch item
73 returns: shape (B, T, hidden_size)
74 """
75 B, T, _ = actions.shape
76
77 # 1) Expand each batch's single scalar time 'tau' across all T steps
78 # so that shape => (B, T)
79 # e.g. if timesteps is (B,), replicate across T
80 if timesteps.dim() == 1 and timesteps.shape[0] == B:
81 # shape (B,) => (B,T)
82 timesteps = timesteps.unsqueeze(1).expand(-1, T)
83 else:
84 raise ValueError(
85 "Expected `timesteps` to have shape (B,) so we can replicate across T."
86 )
87
88 # 2) Standard action MLP step for shape => (B, T, w)
89 a_emb = self.W1(actions)
90
91 # 3) Get the sinusoidal encoding (B, T, w)
92 tau_emb = self.pos_encoding(timesteps).to(dtype=a_emb.dtype)
93
94 # 4) Concat along last dim => (B, T, 2w), then W2 => (B, T, w), swish
95 x = torch.cat([a_emb, tau_emb], dim=-1)
96 x = swish(self.W2(x))
97
98 # 5) Finally W3 => (B, T, w)
99 x = self.W3(x)
100
101 return x

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected