Training code about octconv.pytorch

[d-li14]

An experimental training script is provided here. As it is a trivial implementation resembling the official tutorial, I didn't include it in this repo.

As for hyperparameters, I strictly follow the ResNet paper other than decaying the LR as a cosine function shape during the total 120 epochs.

Transferring to ResNet built with basic block follows the same principle by setting alpha_in to zero in the first stage. Temporarily, it is not the focus of the original paper as well as my reproduction plan.

[gasvn]

Thanks for your excellent work. I am wondering if you could share the "cosine function shape"- adjust_learning_rate function of your training code. Thanks!