search

icon-lab / SynDiff

Official PyTorch implementation of SynDiff described in the paper (https://arxiv.org/abs/2207.08208).
Other
244 stars 37 forks source link

Potential validation/testing leakage due to the absence of switching between train and eval modes #46

Open jinhaoduan opened 4 months ago

jinhaoduan commented 4 months ago

Thanks for releasing the code.

I noticed that there's no switching between model.train() mode and model.eval() mode when transitioning between the training and validation processes. Could this potentially lead to leakage of validation/test information (BN will track the statistics in the validation/testing set)?

Another minor issue is the incorrect parameter in https://github.com/icon-lab/SynDiff/blob/8e5064e6f1a798a166cee5e98c66700bab92b0cc/train.py#L699 The generator should be gen_diffusive_2 for T1 -> T2, right?