| | |
|---|
| | | if cache["cif_alphas"] is not None:
|
|---|
| | | alphas = torch.cat((cache["cif_alphas"], alphas), -1)
|
|---|
| | |
|
|---|
| | | #if cache["is_final"]:
|
|---|
| | | # tail_threshold = torch.tensor([self.tail_threshold], dtype=alphas.dtype).to(alphas.device)
|
|---|
| | | # tail_threshold = torch.reshape(tail_threshold, (1, 1))
|
|---|
| | | # alphas = torch.cat([alphas, tail_threshold], dim=1)
|
|---|
| | | # zeros_hidden = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)
|
|---|
| | | # hidden = torch.cat([hidden, zeros_hidden], dim=1)
|
|---|
| | |
|
|---|
| | | token_num = alphas.sum(-1)
|
|---|
| | | acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)
|
|---|
| | | len_time = alphas.size(-1)
|
|---|
| | | last_fire_place = len_time - 1
|
|---|
| | | last_fire_remainds = 0.0
|
|---|
| | | pre_alphas_length = 0
|
|---|
| | | last_fire = False
|
|---|
| | |
|
|---|
| | | mask_chunk_peak_predictor = None
|
|---|
| | | if cache is not None:
|
|---|
| | |
|---|
| | | pre_alphas_length = cache["cif_alphas"].size(-1)
|
|---|
| | | mask_chunk_peak_predictor[:, :pre_alphas_length] = 1.0
|
|---|
| | | mask_chunk_peak_predictor[:, pre_alphas_length + cache["pad_left"]:pre_alphas_length + cache["stride"] + cache["pad_left"]] = 1.0
|
|---|
| | | #if cache["is_final"]:
|
|---|
| | | # mask_chunk_peak_predictor[:, -1] = 1.0
|
|---|
| | |
|
|---|
| | | if mask_chunk_peak_predictor is not None:
|
|---|
| | | cif_peak = cif_peak * mask_chunk_peak_predictor.squeeze(-1)
|
|---|
| | |
|---|
| | | if cif_peak[0][len_time - 1 - i] > self.threshold or cif_peak[0][len_time - 1 - i] == self.threshold:
|
|---|
| | | last_fire_place = len_time - 1 - i
|
|---|
| | | last_fire_remainds = cif_peak[0][len_time - 1 - i] - self.threshold
|
|---|
| | | last_fire = True
|
|---|
| | | break
|
|---|
| | | last_fire_remainds = torch.tensor([last_fire_remainds], dtype=alphas.dtype).to(alphas.device)
|
|---|
| | | cache["cif_hidden"] = hidden[:, last_fire_place:, :]
|
|---|
| | | cache["cif_alphas"] = torch.cat((last_fire_remainds.unsqueeze(0), alphas[:, last_fire_place+1:]), -1)
|
|---|
| | | if last_fire:
|
|---|
| | | last_fire_remainds = torch.tensor([last_fire_remainds], dtype=alphas.dtype).to(alphas.device)
|
|---|
| | | cache["cif_hidden"] = hidden[:, last_fire_place:, :]
|
|---|
| | | cache["cif_alphas"] = torch.cat((last_fire_remainds.unsqueeze(0), alphas[:, last_fire_place+1:]), -1)
|
|---|
| | | else:
|
|---|
| | | cache["cif_hidden"] = hidden
|
|---|
| | | cache["cif_alphas"] = alphas
|
|---|
| | | token_num_int = token_num.floor().type(torch.int32).item()
|
|---|
| | | return acoustic_embeds[:, 0:token_num_int, :], token_num, alphas, cif_peak
|
|---|
| | |
|
|---|
