Method forward

cdslib/core/nn/modules/vrnn.py:483–574

Run the network T (= seq_len) times and generate T outputs, where T is the sequence length of xs. Args: xs: (seq_len, batch, dim_x) Past input sequence (starting from x0, which is usually zero). Can be constructed by calling construct_tea

(
        self,
        xs: torch.Tensor,
        cs: torch.Tensor,
        fs: torch.Tensor,
        gs: torch.Tensor = None,
        init_hidden_states: T.Dict[str, torch.Tensor] = None,
        valid_lens_fs: T.Sequence[int] = None,
        pz_multiplier: float = 0.0,
        qz_multiplier: float = 1.0,
        pz_std_multiplier: float = 1.0,
        qz_std_multiplier: float = 1.0,
    )

        }

    def forward(
        self,
        xs: torch.Tensor,
        cs: torch.Tensor,
        fs: torch.Tensor,
        gs: torch.Tensor = None,
        init_hidden_states: T.Dict[str, torch.Tensor] = None,
        valid_lens_fs: T.Sequence[int] = None,
        pz_multiplier: float = 0.0,
        qz_multiplier: float = 1.0,
        pz_std_multiplier: float = 1.0,
        qz_std_multiplier: float = 1.0,
    ):
        """
        Run the network T (= seq_len) times and generate T outputs, where T is the sequence length of xs.

        Args:
            xs: (seq_len, batch, dim_x)
                Past input sequence (starting from x0, which is usually zero).
                Can be constructed by calling construct_teacher_vectors(xs_gt, past_to_use=1).
            cs: (num_chars, batch, dim_c)
                Content sequence (can be onehot encoded or some embedding)
            gs: (seq_len, batch, dim_g)
                Side information sequence (gs[t] is concatenated to
                sampled_latent before transformed into z and given to the denoder_rnn)
            fs: (seq_len_2, batch, dim_f)
                Feature sequence that the latent variables will be extracted from
            init_hidden_states:
                A dict containing the hidden states of the model
                'attn_rnn_h'
                'decode_rnn_h'
                'attn_c'
                'attn_mean_idxs'

        Returns:
            A dict containing:
                ps: (seq_len, batch, dim_p)
                attn_cs: (seq_len, batch, dim_c)
                attn_weights: (seq_len, batch, num_chars+1)
                attn_fs_weights: (seq_len, batch, seq_len_2)
                attn_hs: (seq_len, batch, dim_attn_h)
                decode_hs: (seq_len, batch, dim_decode_h)
                hidden_states: A dict containing the hidden states of the model

                    - 'attn_rnn_h'
                    - 'decode_rnn_h'
                    - 'attn_c'
                    - 'attn_mean_idxs'

                .. outputs from zt_fun

                zs: (seq_len, batch, dim_z)
                queries: (seq_len, batch, dim_query)
                sampled_latents: (seq_len, batch, dim_latent)
                pz_means: (seq_len, batch, dim_latent)
                pz_stds: (seq_len, batch, dim_latent)
                qz_means: (seq_len, batch, dim_latent)
                qz_stds: (seq_len, batch, dim_latent)