SOTA Open Source TTS
Optimize training loop & upgrade transformers
Lengyue23109ecb9440
diff --git a/requirements.txt b/requirements.txtindex 5b68898..0fee1db 100644--- a/requirements.txt+++ b/requirements.txt@@ -1,4 +1,4 @@-transformers>=4.34.0+transformers>=4.34.1datasets>=2.14.5bitsandbytes>=0.41.1peft>=0.5.0diff --git a/speech_lm/configs/whisper_vq.yaml b/speech_lm/configs/whisper_vq.yamlindex b2190e8..5dbafaf 100644--- a/speech_lm/configs/whisper_vq.yaml+++ b/speech_lm/configs/whisper_vq.yaml@@ -43,7 +43,7 @@ schedule:save_interval: 2000gradient_accumulation_steps: "${eval: ${schedule.batch_size} // ${schedule.micro_batch_size}}"clip_grad_norm: 2.0- log_interval: 10+ log_interval: 50eval_interval: 2000train_dataloader:@@ -52,7 +52,11 @@ train_dataloader:_target_: speech_lm.datasets.whisper_vq.WhisperVQDatasetfilelist: filelists/whisper-vq.train.filelistbatch_size: ${schedule.micro_batch_size}- num_workers: 8+ num_workers: 16+ prefetch_factor: 4+ pin_memory: true+ persistent_workers: true+ shuffle: truecollate_fn:_target_: speech_lm.datasets.whisper_vq.WhisperVQCollator@@ -62,7 +66,10 @@ valid_dataloader:_target_: speech_lm.datasets.whisper_vq.WhisperVQDatasetfilelist: filelists/whisper-vq.test.filelistbatch_size: 32- num_workers: 4+ num_workers: 8+ prefetch_factor: 4+ pin_memory: true+ shuffle: falsecollate_fn:_target_: speech_lm.datasets.whisper_vq.WhisperVQCollatordiff --git a/speech_lm/train.py b/speech_lm/train.pyindex f2c3982..66a0cef 100644--- a/speech_lm/train.py+++ b/speech_lm/train.py@@ -37,7 +37,7 @@ def valid(accumulate_infos = None- for idx, batch in tqdm(enumerate(valid_dataloader), desc="Evaluating"):+ for idx, batch in enumerate(tqdm(valid_dataloader, desc="Evaluating")):outputs = model(**batch)loss = outputs.lossmetrics = getattr(outputs, "metrics", {})@@ -81,8 +81,6 @@ def train(fabric: Fabric,cfg: DictConfig,):- bar = tqdm(total=cfg.schedule.max_steps, desc="Training")- bar.update(global_step)accumulate_steps = 0optimizer.zero_grad()@@ -163,7 +161,6 @@ def train()global_step += 1- bar.update(1)if global_step % cfg.schedule.log_interval == 0:step_time = (time.time() - start_time) / cfg.schedule.log_interval