I would update PyTorch version to 1.3 / 1.4.
Do we really need a separate TensorboardX package installation? As I know, it's part of PyTorch now (torch.utils.tensorboard).
Maybe it has the sense to delete wandb from the base image and make the installation by a flag in cookiecutter? Also, it can simplify Makefile in this case. It is really annoying to delete wandb files from every inited template if you don't use it.
Do we really need to build CMake?
By @atselousov . origin: https://neuromation.slack.com/archives/CGP770K27/p1582273220132600?thread_ts=1582107299.115400&cid=CGP770K27