What is WandB

WandB is a handy tool to visualize training process as well as cross-validate hyper-parameters. One thing wandb do is it automatically uploads all files in wandb.run.dir. However you only get 100GB for free, and when you save a lot of checkpoints, videos, images, it’s easily become full.

While I know we can just create another folder to save our large files, but I kind of like how wandb name each run, so I still prefer saving to wandb.run.dir. (and sometimes it took too long to upload all files…)

The official doc very briefly talked about how to ignore files for your project, but there is not demo files, I did some research in their source code, and found the steps.

  1. Create a file named setting in wandb/
  2. Edit setting, both lines has to be exactly like this and for globs, each file should be comma separated, no extra space in between. e.g.
[default]
ignore_globs=*.npy,*pt