Training code of pretrain models in torchvision

PyTorch release popular models in torchvision with pre-trained checkpoints.

I want to reproduce the training process of these models, e.g. mobilenetv2, are there any official training code for those pre-trained models?

I see there is an example to train networks with ImageNet, but I don’t know how to set these arguments for different networks.

You can define different models by passing the --arch argument with the desired model name.
These lines of code will then create the model (and state_dict, if you want to use a pretrained model).

But how can I know the other parameters used to train the network?
For example, as mentioned in the MobileNet v2 paper, they train their network with RMSProp and some special learning rate schedule. Some implementations use extra data augmentation (e.g. label smoothing). I want to know about these training details when train the pretrain network.