(self, txt_tokens, mel2ph=None, spk_embed=None,
ref_mels=None, f0=None, uv=None, energy=None, infer=False)
| 311 | return loss |
| 312 | |
| 313 | def forward(self, txt_tokens, mel2ph=None, spk_embed=None, |
| 314 | ref_mels=None, f0=None, uv=None, energy=None, infer=False): |
| 315 | b, *_, device = *txt_tokens.shape, txt_tokens.device |
| 316 | ret = self.fs2(txt_tokens, mel2ph, spk_embed, ref_mels, f0, uv, energy, |
| 317 | skip_decoder=(not infer), infer=infer) |
| 318 | cond = ret['decoder_inp'].transpose(1, 2) |
| 319 | |
| 320 | if not infer: |
| 321 | t = torch.randint(0, self.K_step, (b,), device=device).long() |
| 322 | x = ref_mels |
| 323 | x = self.norm_spec(x) |
| 324 | x = x.transpose(1, 2)[:, None, :, :] # [B, 1, M, T] |
| 325 | ret['diff_loss'] = self.p_losses(x, t, cond) |
| 326 | # nonpadding = (mel2ph != 0).float() |
| 327 | # ret['diff_loss'] = self.p_losses(x, t, cond, nonpadding=nonpadding) |
| 328 | else: |
| 329 | ret['fs2_mel'] = ret['mel_out'] |
| 330 | fs2_mels = ret['mel_out'] |
| 331 | t = self.K_step |
| 332 | fs2_mels = self.norm_spec(fs2_mels) |
| 333 | fs2_mels = fs2_mels.transpose(1, 2)[:, None, :, :] |
| 334 | |
| 335 | x = self.q_sample(x_start=fs2_mels, t=torch.tensor([t - 1], device=device).long()) |
| 336 | for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t): |
| 337 | x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond) |
| 338 | x = x[:, 0].transpose(1, 2) |
| 339 | ret['mel_out'] = self.denorm_spec(x) |
| 340 | |
| 341 | return ret |
| 342 | |
| 343 | def norm_spec(self, x): |
| 344 | return (x - self.spec_min) / (self.spec_max - self.spec_min) * 2 - 1 |
nothing calls this directly
no test coverage detected