(
self,
num_mode,
decoder_depth,
dim,
num_heads,
mlp_ratio,
dropout,
future_steps,
yaw_constraint=False,
cat_x=False,
)
| 88 | |
| 89 | class PlanningDecoder(nn.Module): |
| 90 | def __init__( |
| 91 | self, |
| 92 | num_mode, |
| 93 | decoder_depth, |
| 94 | dim, |
| 95 | num_heads, |
| 96 | mlp_ratio, |
| 97 | dropout, |
| 98 | future_steps, |
| 99 | yaw_constraint=False, |
| 100 | cat_x=False, |
| 101 | ) -> None: |
| 102 | super().__init__() |
| 103 | |
| 104 | self.num_mode = num_mode |
| 105 | self.future_steps = future_steps |
| 106 | self.yaw_constraint = yaw_constraint |
| 107 | self.cat_x = cat_x |
| 108 | |
| 109 | self.decoder_blocks = nn.ModuleList( |
| 110 | [ |
| 111 | DecoderLayer(dim, num_heads, mlp_ratio, dropout) |
| 112 | for _ in range(decoder_depth) |
| 113 | ] |
| 114 | ) |
| 115 | |
| 116 | self.r_pos_emb = FourierEmbedding(3, dim, 64) |
| 117 | self.r_encoder = PointsEncoder(6, dim) |
| 118 | |
| 119 | self.q_proj = nn.Linear(2 * dim, dim) |
| 120 | |
| 121 | self.m_emb = nn.Parameter(torch.Tensor(1, 1, num_mode, dim)) |
| 122 | self.m_pos = nn.Parameter(torch.Tensor(1, num_mode, dim)) |
| 123 | |
| 124 | if self.cat_x: |
| 125 | self.cat_x_proj = nn.Linear(2 * dim, dim) |
| 126 | |
| 127 | self.loc_head = MLPLayer(dim, 2 * dim, self.future_steps * 2) |
| 128 | self.yaw_head = MLPLayer(dim, 2 * dim, self.future_steps * 2) |
| 129 | self.vel_head = MLPLayer(dim, 2 * dim, self.future_steps * 2) |
| 130 | self.pi_head = MLPLayer(dim, dim, 1) |
| 131 | |
| 132 | nn.init.normal_(self.m_emb, mean=0.0, std=0.01) |
| 133 | nn.init.normal_(self.m_pos, mean=0.0, std=0.01) |
| 134 | |
| 135 | def forward(self, data, enc_data): |
| 136 | enc_emb = enc_data["enc_emb"] |
nothing calls this directly
no test coverage detected