| 282 | ) |
| 283 | |
| 284 | def score(self, yseq, state, x): |
| 285 | # to support mutiple encoder asr mode, in single encoder mode, |
| 286 | # convert torch.Tensor to List of torch.Tensor |
| 287 | if self.num_encs == 1: |
| 288 | x = [x] |
| 289 | |
| 290 | att_idx, z_list, c_list = state["workspace"] |
| 291 | vy = yseq[-1].unsqueeze(0) |
| 292 | ey = self.dropout_emb(self.embed(vy)) # utt list (1) x zdim |
| 293 | if self.num_encs == 1: |
| 294 | att_c, att_w = self.att_list[att_idx]( |
| 295 | x[0].unsqueeze(0), |
| 296 | [x[0].size(0)], |
| 297 | self.dropout_dec[0](state["z_prev"][0]), |
| 298 | state["a_prev"], |
| 299 | ) |
| 300 | else: |
| 301 | att_w = [None] * (self.num_encs + 1) # atts + han |
| 302 | att_c_list = [None] * self.num_encs # atts |
| 303 | for idx in range(self.num_encs): |
| 304 | att_c_list[idx], att_w[idx] = self.att_list[idx]( |
| 305 | x[idx].unsqueeze(0), |
| 306 | [x[idx].size(0)], |
| 307 | self.dropout_dec[0](state["z_prev"][0]), |
| 308 | state["a_prev"][idx], |
| 309 | ) |
| 310 | h_han = torch.stack(att_c_list, dim=1) |
| 311 | att_c, att_w[self.num_encs] = self.att_list[self.num_encs]( |
| 312 | h_han, |
| 313 | [self.num_encs], |
| 314 | self.dropout_dec[0](state["z_prev"][0]), |
| 315 | state["a_prev"][self.num_encs], |
| 316 | ) |
| 317 | ey = torch.cat((ey, att_c), dim=1) # utt(1) x (zdim + hdim) |
| 318 | z_list, c_list = self.rnn_forward( |
| 319 | ey, z_list, c_list, state["z_prev"], state["c_prev"] |
| 320 | ) |
| 321 | if self.context_residual: |
| 322 | logits = self.output( |
| 323 | torch.cat((self.dropout_dec[-1](z_list[-1]), att_c), dim=-1) |
| 324 | ) |
| 325 | else: |
| 326 | logits = self.output(self.dropout_dec[-1](z_list[-1])) |
| 327 | logp = F.log_softmax(logits, dim=1).squeeze(0) |
| 328 | return ( |
| 329 | logp, |
| 330 | dict( |
| 331 | c_prev=c_list[:], |
| 332 | z_prev=z_list[:], |
| 333 | a_prev=att_w, |
| 334 | workspace=(att_idx, z_list, c_list), |
| 335 | ), |
| 336 | ) |