Training code

[miguelvr]

Are you planning on releasing the training code?

Also, did you try to implement the ResNet BasicBlock with OctConv?

I'm trying to do it in my own implementation, but it is tricky due to the lack of down sampling on the first layer.

[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!