You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Suggestion: change the default value of n_signal in train.py to 122880.
If I'm not mistaken, the current defaults for preprocess.py (train.py) set num_signal (n_signal) to the same value, 131072, which means the random cropping data augmentation is effectively turned off by default. I'm also seeing occasional OOM when training v3 models on a 16 GB GPU. To address both, I made n_signal in train.py slightly shorter (122880).
Posting this as an issue since it might be more user-friendly to adopt as the default.
The text was updated successfully, but these errors were encountered:
Suggestion: change the default value of
n_signal
intrain.py
to 122880.If I'm not mistaken, the current defaults for
preprocess.py
(train.py
) setnum_signal
(n_signal
) to the same value, 131072, which means the random cropping data augmentation is effectively turned off by default. I'm also seeing occasional OOM when training v3 models on a 16 GB GPU. To address both, I maden_signal
intrain.py
slightly shorter (122880).Posting this as an issue since it might be more user-friendly to adopt as the default.
The text was updated successfully, but these errors were encountered: