73.0% Top-1 with 200 epochs training without color jittering

[cjnjuwhy]

Hi, I tried to train the MobileNeXt model in ImageNet dataset, in short, the best Top-1 is 73.0% at the 196th epoch of 200. For fast training, I set the 2x batch_size and 2x initial LR, there is no color jittering in training time and other details are the same with your paper, more details please refer to train.log in my repo.

However, there is a small question, the MobileNeXt structure modified by myself are 319M MultAdds, which is a little different with 300M reported in the paper. I guess something is different in the hidden layer, here is my model definition.