SOTA Open Source TTS
Fix dataset and training loop
Lengyue01cf9bbafd4c
diff --git a/speech_lm/datasets/cultura_x.py b/speech_lm/datasets/cultura_x.pyindex fa564ee..9445000 100644--- a/speech_lm/datasets/cultura_x.py+++ b/speech_lm/datasets/cultura_x.py@@ -67,8 +67,8 @@ class CulturaXDataset(IterableDataset):for filename in files:try:yield from self.parse_data(filename)- except:- log.exception(f"Failed to parse {filename}")+ except Exception as e:+ log.exception(f"Failed to parse {filename}: {e}")def parse_data(self, filename: str):url = f"https://huggingface.co/datasets/uonlp/CulturaX/resolve/main/{filename}"diff --git a/speech_lm/train.py b/speech_lm/train.pyindex ca82941..5b68f43 100644--- a/speech_lm/train.py+++ b/speech_lm/train.py@@ -117,6 +117,9 @@ def main(cfg: DictConfig):log.info(f"Optimizer: {optimizer}")log.info(f"Scheduler: {scheduler}")+ log.info(f"Setup fabric model & dataset")+ model, optimizer, scheduler = fabric.setup(model, optimizer, scheduler)+# Build stateglobal_step = 0@@ -141,9 +144,6 @@ def main(cfg: DictConfig):global_step = remainder["global_step"]log.info(f"Restored global step: {global_step}")- log.info(f"Setup fabric model & dataset")- model, optimizer, scheduler = fabric.setup(model, optimizer, scheduler)-train_dataloader = hydra.utils.instantiate(cfg.dataloader)log.info(f"Dataloader: {train_dataloader}")