mrq/vall-e
5
9
Fork 0
Code Issues 8 Pull Requests Packages Projects Releases Wiki Activity

remove double spaces in the text phonemes (might have caused problems.........)

Browse Source
This commit is contained in:
mrq 2023-10-10 19:18:24 -05:00
parent 47b3077415
commit b4405c98ea
1 changed files with 12 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
vall_e/data.py
View File

@ -143,9 +143,14 @@ def _get_paths_of_extensions( path, extensions=".qnt.pt", validate=False ):
def _load_quants(path) -> Tensor: def _load_quants(path) -> Tensor:
return torch.load(_get_quant_path(path))[0][:, :].t().to(torch.int16) return torch.load(_get_quant_path(path))[0][:, :].t().to(torch.int16)
# prune consecutive spaces
def _cleanup_phones( phones, targets=[" "]):
return [ p for i, p in enumerate(phones) if p not in targets or ( p in targets and p != phones[i-1] ) ]
@cache @cache
def _get_phones(path, language="en"): def _get_phones(path, language="en"):
content = open(_get_phone_path(path), "r", encoding="utf-8").read().split(" ") content = open(_get_phone_path(path), "r", encoding="utf-8").read().split(" ")
content = _cleanup_phones( content )
return ["<s>"] + [ " " if not p else p for p in content ] + ["</s>"] return ["<s>"] + [ " " if not p else p for p in content ] + ["</s>"]
def _interleaved_reorder(l, fn): def _interleaved_reorder(l, fn):
@ -333,8 +338,13 @@ class Dataset(_Dataset):
if cfg.dataset.use_hdf5: if cfg.dataset.use_hdf5:
key = _get_hdf5_path(path) key = _get_hdf5_path(path)
text = torch.from_numpy(cfg.hdf5[key]["text"][:]).to(self.text_dtype) text = cfg.hdf5[key]["text"][:]
resps = torch.from_numpy(cfg.hdf5[key]["audio"][:, :]).to(torch.int16) resps = cfg.hdf5[key]["audio"][:, :]
text = np.array( _cleanup_phones( text, targets=[ self.phone_symmap[" "] ] ) )
text = torch.from_numpy(text).to(self.text_dtype)
resps = torch.from_numpy(resps).to(torch.int16)
else: else:
text = torch.tensor([*map(self.phone_symmap.get, _get_phones(path))]).to(self.text_dtype) text = torch.tensor([*map(self.phone_symmap.get, _get_phones(path))]).to(self.text_dtype)
resps = _load_quants(path) resps = _load_quants(path)