search

gan-police / frequency-forensics

Deepfake detection using wavelet-packets in PyTorch, European Conference on Machine Learning (ECML PKDD) 2022.
Other
47 stars 9 forks source link

No Wavelet decomposition during the training #3

Closed felixblanke closed 3 years ago

felixblanke commented 3 years ago

It appears to me that we are always training on the raw normalized images and calculate the wavelet representation only in the validation loop. Am I missing something?

https://github.com/gan-police/frequency-forensics/blob/149db532c0c1f918a983eb60405ff4f69ff6b94d/src/freqdect/train_classifier.py#L104-L108

https://github.com/gan-police/frequency-forensics/blob/149db532c0c1f918a983eb60405ff4f69ff6b94d/src/freqdect/train_classifier.py#L31-L43

v0lta commented 3 years ago

ahh yeah, I broke this sorry!

v0lta commented 3 years ago

I am working on batched preprocessing to compute the wavelet packets efficiently ahead of the training.

v0lta commented 3 years ago

Until that is ready we need it in the training code. I should not have removed it, sorry.

v0lta commented 3 years ago

packet_raw_ffhw_stylegan I fixed it. Results on the FFHQ - Stylegan pair look promising.