search

seongmin-kye / MCT

Pytorch implementation of Meta-Learned Confidence for Few-shot Learning
71 stars 12 forks source link

issue about the preprocessing #12

Closed CSer-Tang-hao closed 4 years ago

CSer-Tang-hao commented 4 years ago

I noticed that the operation similar to "transforms. Normalize" is not used in data preprocessing, but this is a common data processing step that used in most previous works such as MetaOptNet.

Looking forward to your reply!

Thanks

seongmin-kye commented 4 years ago

I think that normalization step is not suitable for practical scenarios, since training data and classes can be incremental. But, normalization with mean & std should know data statistics.

Thanks.

CSer-Tang-hao commented 4 years ago

I think that normalization step is not suitable for practical scenarios, since training data and classes can be incremental. But, normalization with mean & std should know data statistics.

Thanks.

In terms of the use of the normalization step, have you done any experiments to show that it's affecting?

seongmin-kye commented 4 years ago

I haven't experimented separately. It would be great if you could try and let me know.