what's the L1Loss value when pretrain model's training finished ? about s3prl

[andi611]

There is no exact value, it largely depends on the feature (or preprocessing setting) you used.

[apoca909]

There is no exact value, it largely depends on the feature (or preprocessing setting) you used.

I use the default args setting in extract_feature func.

feature='fbank',
delta=False,
delta_delta=False,
cmvn=True

[andi611]

For fbank, it should be within the range of 0.85~0.95.
Don't worry too much about the L1Loss value.

FYI, we recommend you pre-train with our latest feature: on-the-fly preprocessing.
This omits data preprocessing beforehand, and the trained model will have a built-in preprocessor that can take wave as input.

[apoca909]

For fbank, it should be within the range of 0.85~0.95.
Don't worry too much about the L1Loss value.

FYI, we recommend you pre-train with our latest feature: on-the-fly preprocessing.
This omits data preprocessing beforehand, and the trained model will have a built-in preprocessor that can take wave as input.

very much thx.