668c3ef362995c55633fde592354160fec1d1efd,onmt/decoders/transformer.py,TransformerDecoder,forward,#TransformerDecoder#Any#Any#Any#,189

Before Change

            self._init_cache(memory_bank)

        src = self.state["src"]
        src_words = src[:, :, 0].transpose(0, 1)
        tgt_words = tgt[:, :, 0].transpose(0, 1)
        src_batch, src_len = src_words.size()
        tgt_batch, tgt_len = tgt_words.size()

        emb = self.embeddings(tgt, step=step)
        assert emb.dim() == 3  // len x batch x embedding_dim

After Change

        pad_idx = self.embeddings.word_padding_idx
        src_lens = kwargs["memory_lengths"]
        src_max_len = self.state["src"].shape[0]
        src_pad_mask = ~sequence_mask(src_lens, src_max_len).unsqueeze(1)
        tgt_pad_mask = tgt_words.data.eq(pad_idx).unsqueeze(1)  // [B, 1, T_tgt]

        for i, layer in enumerate(self.transformer_layers):
            layer_cache = self.state["cache"]["layer_{}".format(i)] \

Instances