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

add config option to set the rate of sampling randomly vs similar speakers during training

Browse Source
This commit is contained in:
mrq 2024-10-16 14:27:58 -05:00
parent 48461833c2
commit f88097ccf6
2 changed files with 4 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
vall_e/config.py
View File

@ -158,8 +158,9 @@ class Dataset:
max_resps: int = 1 # number of samples to target for training
p_resp_append: float = 1.0 # probability to append another sample to the training target
p_resp_pad_silence: float = 0.0 # probability to pad resp with silence to fit within the next window
prompt_similar_top_k: int = 1
prompt_similar_top_k_offset: int = 0
prompt_similar_p: float = 0.75 # odds of sampling for a similar prompt instead of a random prompt
prompt_similar_top_k: int = 1 # top-k similar candidates to sample from
prompt_similar_top_k_offset: int = 0 # offset from the top-k to sample from
sample_type: str = "path" # path | speaker
sample_order: str = "interleaved" # duration

2
vall_e/data.py
View File

@ -1014,7 +1014,7 @@ class Dataset(_Dataset):
for _ in range(cfg.dataset.max_prompts):
if reference is not None and cfg.dataset.prom_sample_similar:
path = self.get_similar_utterance( reference, offset = len(prom_list) )
path = self.get_similar_utterance( reference, offset = len(prom_list) ) if random.random() < cfg.dataset.prompt_similar_p else random.choice(choices)
# yuck
if not path:
path = random.choice(choices)