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

added option to retokenize phonemes for hdf5 (to save having to remake my hdf5 file)

Browse Source
This commit is contained in:
mrq 2024-09-21 13:08:01 -05:00
parent 536c11c4ac
commit c5e9142863
2 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vall_e/config.py
View File

@ -173,7 +173,8 @@ class Dataset:
reencode_device: str = "cpu" # "cpu" is slower but saves memory, cuda throws [rank0]: RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
noise_scale: float = 0.25 # scaling noise value
inject_noise_in_prom: bool = False # adds noise to the input prompt waveform to try and vary things
retokenize_text: bool = False
_frames_per_second: int = 0 # allows setting your own hint
@cached_property

3
vall_e/data.py
View File

@ -1005,6 +1005,9 @@ class Dataset(_Dataset):
lang = metadata["language"] if "language" in metadata else None
tone = metadata["tone"] if "tone" in metadata else None
text_string = metadata["text"] if "text" in metadata else None
if cfg.dataset.retokenize_text and "phonemes" in metadata:
text = torch.tensor(tokenize( metadata["phonemes"] )).to(self.text_dtype)
else:
resps, metadata = _load_quants(path, return_metadata=True)
text = torch.tensor(tokenize( metadata["phonemes"] )).to(self.text_dtype)