Args: dec_inputs: [batch_size, tgt_len] enc_inputs: [batch_size, src_len] enc_outputs: [batch_size, src_len, d_model]
(self, dec_inputs, enc_inputs, enc_outputs)
| 128 | self.layers.append(TransformerDecoderLayer(d_model=d_model, n_head=n_head, dim_feedforward=dim_feedforward)) |
| 129 | |
| 130 | def forward(self, dec_inputs, enc_inputs, enc_outputs): |
| 131 | """ |
| 132 | Args: |
| 133 | dec_inputs: [batch_size, tgt_len] |
| 134 | enc_inputs: [batch_size, src_len] |
| 135 | enc_outputs: [batch_size, src_len, d_model] |
| 136 | |
| 137 | """ |
| 138 | |
| 139 | # [batch_size, tgt_len, d_model] |
| 140 | tgt_word_emb = self.target_emb(dec_inputs) |
| 141 | self.target_pos_emb.initialize(dec_inputs) |
| 142 | self.target_pos_emb.from_pretrained(W=TransformerDecoder._get_sinusoid_encoding_table(self.tgt_n_token, self.d_model), |
| 143 | freeze=True) |
| 144 | # [batch_size, tgt_len, d_model] |
| 145 | tgt_pos_emb = self.target_pos_emb(dec_inputs) |
| 146 | # [batch_size, tgt_len, d_model] |
| 147 | dec_outputs = autograd.add(tgt_word_emb, tgt_pos_emb) |
| 148 | |
| 149 | # dec_self_attn_pad_mask [batch_size, tgt_len, tgt_len] |
| 150 | dec_self_attn_pad_mask = TransformerDecoder._get_attn_pad_mask(dec_inputs, dec_inputs) |
| 151 | # [batch_size, tgt_len, tgt_len] |
| 152 | dec_self_attn_subsequent_mask = TransformerDecoder._get_attn_subsequence_mask(dec_inputs) |
| 153 | |
| 154 | # dec_self_attn_mask [batch_size, tgt_len, tgt_len] |
| 155 | dec_self_attn_mask = tensor.gt((dec_self_attn_pad_mask + dec_self_attn_subsequent_mask), 0) |
| 156 | |
| 157 | # dec_enc_attn_mask [batch_size, tgt_len, src_len] |
| 158 | dec_enc_attn_mask = TransformerDecoder._get_attn_pad_mask(dec_inputs, enc_inputs) |
| 159 | |
| 160 | dec_self_attns, dec_enc_attns = [], [] |
| 161 | |
| 162 | for layer in self.layers: |
| 163 | # dec_outputs: [batch_size, tgt_len, d_model], |
| 164 | # dec_self_attn: [batch_size, n_heads, tgt_len, tgt_len], |
| 165 | # dec_enc_attn: [batch_size, h_heads, tgt_len,src_len] |
| 166 | dec_outputs, dec_self_attn, dec_enc_attn = layer(dec_outputs, enc_outputs, dec_self_attn_mask, |
| 167 | dec_enc_attn_mask) |
| 168 | dec_self_attns.append(dec_self_attn) |
| 169 | dec_enc_attns.append(dec_enc_attn) |
| 170 | return dec_outputs, dec_self_attns, dec_enc_attns |
| 171 | |
| 172 | @staticmethod |
| 173 | def _get_attn_pad_mask(seq_q, seq_k): |
nothing calls this directly
no test coverage detected