blakelemoine/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

better bounds

Browse Source
This commit is contained in:
James Betker 2022-01-01 14:05:22 -07:00
parent fe9ea4e01a
commit dc535b5358
1 changed files with 7 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
codes/models/gpt_voice/gpt_asr_hf2.py
View File

@ -288,6 +288,7 @@ class GptAsrHf2(nn.Module):
mel_len = 0 mel_len = 0
else: else:
mel_emb = self.mel_encoder(mel_inputs) mel_emb = self.mel_encoder(mel_inputs)
assert mel_emb.shape[1] <= self.max_mel_frames
mel_emb = mel_emb.permute(0,2,1).contiguous() mel_emb = mel_emb.permute(0,2,1).contiguous()
mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device)) mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
emb = torch.cat([mel_emb, text_emb], dim=1) emb = torch.cat([mel_emb, text_emb], dim=1)
@ -302,6 +303,9 @@ class GptAsrHf2(nn.Module):
return text_logits return text_logits
def forward(self, mel_inputs, text_inputs, return_attentions=False): def forward(self, mel_inputs, text_inputs, return_attentions=False):
assert text_inputs.shape[1] <= self.max_symbols_per_phrase
assert text_inputs.max() <= self.number_text_tokens
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_token, self.stop_token) text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_token, self.stop_token)
text_emb = self.gpt.get_input_embeddings()(text_inputs) + \ text_emb = self.gpt.get_input_embeddings()(text_inputs) + \
self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device)) self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device))
@ -313,9 +317,9 @@ class GptAsrHf2(nn.Module):
return loss_text.mean(), text_logits return loss_text.mean(), text_logits
def text_only(self, text_inputs): def text_only(self, text_inputs):
if text_inputs.shape[1] > self.max_symbols_per_phrase: assert text_inputs.shape[1] <= self.max_symbols_per_phrase
print(f"Embedding error, provided text_inputs with shape {text_inputs.shape}, but max is {self.max_symbols_per_phrase}. Automatically correcting by truncating symbols.") assert text_inputs.max() <= self.number_text_tokens
text_inputs = text_inputs[:, :self.max_symbols_per_phrase]
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_token, self.stop_token) text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_token, self.stop_token)
text_emb = self.gpt.get_input_embeddings()(text_inputs) + \ text_emb = self.gpt.get_input_embeddings()(text_inputs) + \
self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device)) + \ self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device)) + \