Replies: 1 comment
-
It is possible that the audio file is corrupt (0 length audio). The errors occurs the moment the dataloader tries to read the audio file, so please check for any corrupt audio files. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
When attempting to train a quartznet model on our data we get the error below. The trainer is able to initialize the model but it errors out right as the first epoch begins.
This is the code we are using to train the model
import nemo.collections.asr as nemo_asr
import pytorch_lightning as pl
import torch
import gc
from ruamel.yaml import YAML
from omegaconf import DictConfig
config_path = 'config/config.yaml'
train_manifest = 'utils/manifests/atcc_train.json'
valid_manifest = 'utils/manifests/atcc_validation.json'
checkpoint_name = '100epoch_quartznet_scratch.nemo'
torch.cuda.empty_cache()
gc.collect()
with open(config_path, 'r') as f:
config = YAML(typ='safe').load(f)
config['model']['train_ds']['manifest_filepath'] = train_manifest
config['model']['train_ds']['batch_size'] = 8
config['model']['validation_ds']['manifest_filepath'] = valid_manifest
config['model']['validation_ds']['batch_size'] = 8
model = nemo_asr.models.EncDecCTCModel(cfg=DictConfig(config['model']))
model.setup_training_data(DictConfig(config['model']['train_ds']))
model.setup_validation_data(DictConfig(config['model']['validation_ds']))
trainer = pl.Trainer(gpus=[0], max_epochs=100)
trainer.fit(model)
model.save_to(checkpoint_name)
We are using librosa version 0.9.2.
Beta Was this translation helpful? Give feedback.
All reactions