model.train() instead of model.eval()? Or I'm missing something!

[AliLotfi92]

https://github.com/NVlabs/CSG/blob/175ecdf748fcaf7c84de7b1c2578dbc01821c483/train.py#L211

[chenwydj]

Hi @AliLotfi92,

Thanks for your interest in our work!

We use model.eval() because we want to keep the BN statistics learned in the pretrained checkpoint (e.g. from ImageNet) remain intact during the synthetic training. In other words, we want to avoid the BN from being affected by the features on the synthetic images.

[AliLotfi92]

thanks for your response. That makes much more sense now. For train_seq.py you have this: https://github.com/NVlabs/CSG/blob/175ecdf748fcaf7c84de7b1c2578dbc01821c483/train_seg.py#L193 shouldn't it be the same for the classification task too?

[chenwydj]

@AliLotfi92 The fc_new is learning features from the new domain. It is randomly initialized, instead of having a pretrained checkpoint to load from.