Method __init__

openrec/modeling/decoders/bus_decoder.py:15–69

(self,
                 in_channels,
                 out_channels,
                 nhead=8,
                 num_layers=4,
                 dim_feedforward=2048,
                 dropout=0.1,
                 max_length=25,
                 ignore_index=100,
                 pretraining=False,
                 detach=True)

class BUSDecoder(nn.Module):

    def __init__(self,
                 in_channels,
                 out_channels,
                 nhead=8,
                 num_layers=4,
                 dim_feedforward=2048,
                 dropout=0.1,
                 max_length=25,
                 ignore_index=100,
                 pretraining=False,
                 detach=True):
        super().__init__()
        d_model = in_channels
        self.ignore_index = ignore_index
        self.pretraining = pretraining
        self.d_model = d_model
        self.detach = detach
        self.max_length = max_length + 1  # additional stop token
        self.out_channels = out_channels
        # --------------------------------------------------------------------------
        # decoder specifics
        self.proj = nn.Linear(out_channels, d_model, False)
        self.token_encoder = PositionalEncoding(dropout=0.1,
                                                dim=d_model,
                                                max_len=self.max_length)
        self.pos_encoder = PositionalEncoding(dropout=0.1,
                                              dim=d_model,
                                              max_len=self.max_length)

        self.decoder = nn.ModuleList([
            TransformerBlock(
                d_model=d_model,
                nhead=nhead,
                dim_feedforward=dim_feedforward,
                attention_dropout_rate=dropout,
                residual_dropout_rate=dropout,
                with_self_attn=False,
                with_cross_attn=True,
            ) for i in range(num_layers)
        ])

        v_mask = torch.empty((1, 1, d_model))
        l_mask = torch.empty((1, 1, d_model))
        self.v_mask = nn.Parameter(v_mask)
        self.l_mask = nn.Parameter(l_mask)
        torch.nn.init.uniform_(self.v_mask, -0.001, 0.001)
        torch.nn.init.uniform_(self.l_mask, -0.001, 0.001)

        v_embeding = torch.empty((1, 1, d_model))
        l_embeding = torch.empty((1, 1, d_model))
        self.v_embeding = nn.Parameter(v_embeding)
        self.l_embeding = nn.Parameter(l_embeding)
        torch.nn.init.uniform_(self.v_embeding, -0.001, 0.001)
        torch.nn.init.uniform_(self.l_embeding, -0.001, 0.001)
        self.cls = nn.Linear(d_model, out_channels)

    def forward_decoder(self, q, x, mask=None):
        for decoder_layer in self.decoder: