about train and inference

[lichunxiang93]

should i close switch of weight standardization when inference?

[joe-siyuan-qiao]

Thanks for your interest in WS.

WS uses the same code for training and inference as we have provided in this repo, so you do not need to worry about switching on/off WS for inference. Our implementation of imagenet classification is a good reference for how to use WS in training and inference. The increase in computation time or memory usage brought by WS is also neglectable, so we recommend using our provided codes. But if you'd like to change WS to non-WS convolution for inference, you can do that by replacing the weights (in Conv with WS) with the standardized weights and do non-WS Conv computation. However, this may cause problems if you need to fine-tune the models (or continue training) in the future.

Thanks.