Dai Tran

From what I understand, you can use the trained model from orginal repo for step 1. https://github.com/megvii-model/DetNAS

I think you still need to train step 1 with ImageNet ILSVRC2012_img_train data. From step 1 pretrained model then step 2 can be trained.

you can define by creating more \def and \pic. Read and understand the example code first, then you can modify whatever you want. Good luck

Try this: `evaluation = dict(interval=1000, metric='bbox', type='SubModulesDistEvalHook', classwise = True)` If you like to examine the AP of all classes during the evaluation process.

Hi @owaisCS , I have same question. Did you convert other models from [https://github.com/google-research/vision_transformer](https://github.com/google-research/vision_transformer) to mmaction2's type? I tried to convert and read the keys from the original models, but...