Method __init__

model/layers.py:296–329

(self, args)

class Decoder(nn.Module):
    def __init__(self, args) -> None:
        super().__init__()
        '''
        c1 :128, 120, 120
        c2 :256, 60, 60
        c3 :512, 30, 30
        c4 :1024, 15 ,15
        '''
        token_dim = args.token_dim
        self.tokens = nn.Embedding(args.num_token, token_dim)
        trunc_normal_(self.tokens.weight, std=0.02)

        dims = [1024, 512, 256, 128]
        pe_shapes = [30, 60, 120]

        self.layers = []
        for pe_shape in pe_shapes:
            self.layers.append(LoadLayer(token_dim, drop=.1, bias=False, pe_shape=pe_shape))
        self.cgattention1 = CGAttention(token_dim=token_dim,
                            vis_dim=token_dim,
                            hidden_dim=token_dim,
                            drop=.1,
                            bias=True)
        self.cgattention2 = CGAttention(token_dim=token_dim,
                    vis_dim=token_dim,
                    hidden_dim=token_dim,
                    drop=.1,
                    bias=True)
        self.layers = nn.ModuleList(self.layers)
        self.fuses = []
        for dim in [dims[0], dims[2], dims[3]]:
            self.fuses.append(Fusion(dim, token_dim, token_dim, bias=True))
        self.fuses = nn.ModuleList(self.fuses)
        self.proj = DProjector(text_dim=token_dim, in_dim=token_dim)

    def forward(self, vis, text, pad_mask):
        x_c4, x_c3, x_c2, x_c1 = vis