takuoko

https://github.com/open-mmlab/mmclassification/commit/877ea3080a1509e0fedd80dd954f23704a3211f4 Perhaps in the next version of the release, but the PR here will resolve the issue.

However, I think it is slightly inconvenient that the path of site-packages changes depending on the environment.

@zzc98 @Ezra-Yu Thank you for creating useful components! I have other arcface implementation and I prefer it. Please check it too. https://github.com/open-mmlab/mmclassification/pull/933

I think gt_labels also be put in out_items. What do you think?

You can use ImagNet pretrain models. Check this. https://github.com/open-mmlab/mmclassification/blob/master/configs/resnet/resnet50_8xb8_cub.py#L7-L13

I am a little confused as to whether I should add this PR or not. It is a simple implementation, but I don't want to complicate the implementation by adding...

You are correct. However, the yolox hook is used in pairs with a multi image dataset wrapper. Specifically, the following part is required. https://github.com/open-mmlab/mmdetection/blob/master/mmdet/datasets/dataset_wrappers.py#L397-L415 It might be nice to have...

>What about using 'switch' instead of 'stop', which we can change from pipeline=[A, B, C, D] to pipeline=[A, None, C_, D], which has the effect of 'stop'. That is certainly...

I think train_aug is fine with that. I made a mock with the latest push. Especially for data_aug, I have left the stop function in the implementation, how do you...

SwitchDataAugHook is a little dirty for dataset wrappers. Wish we had a better solution, but...