a-One-Fan/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

more debugging

Browse Source
This commit is contained in:
James Betker 2022-01-01 14:25:27 -07:00
parent d8111e0477
commit d4a6298658
2 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
codes/data/__init__.py
View File

@ -89,7 +89,7 @@ def create_dataset(dataset_opt, return_collate=False):
elif mode == 'grand_conjoined_voice': elif mode == 'grand_conjoined_voice':
from data.audio.grand_conjoined_dataset import GrandConjoinedDataset as D from data.audio.grand_conjoined_dataset import GrandConjoinedDataset as D
from data.zero_pad_dict_collate import ZeroPadDictCollate as C from data.zero_pad_dict_collate import ZeroPadDictCollate as C
if opt_get(dataset_opt, ['needs_collate'], True): if opt_get(dataset_opt, ['needs_collate'], False):
collate = C() collate = C()
else: else:
raise NotImplementedError('Dataset [{:s}] is not recognized.'.format(mode)) raise NotImplementedError('Dataset [{:s}] is not recognized.'.format(mode))

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

@ -288,7 +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 assert mel_emb.shape[1] <= self.max_mel_frames, f'{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)
@ -303,8 +303,8 @@ 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.shape[1] <= self.max_symbols_per_phrase, str(text_inputs.shape[1])
assert text_inputs.max() <= self.number_text_tokens assert text_inputs.max() <= self.number_text_tokens, str(text_inputs.max())
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) + \
@ -317,8 +317,8 @@ 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):
assert text_inputs.shape[1] <= self.max_symbols_per_phrase assert text_inputs.shape[1] <= self.max_symbols_per_phrase, str(text_inputs.shape[1])
assert text_inputs.max() <= self.number_text_tokens assert text_inputs.max() <= self.number_text_tokens, str(text_inputs.max())
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) + \