Weights and Biases (wandb)

Easy-to-use, structured experiment versioning tracker for ML projects. Quite simple to orchestrate with PyTorch. Here are some general notes.

• The main structure of wandb is composed of project and runs. You create a project and everytime you e.g., train your model, create a new run. This will let you view all your runs at once on the web server.

wand.init

Lots of what you specify here is what is relevant to your run. Some common things I like to do (view all options here):

Some notes:

• entity is user or org to where your runs are sent. If you have multiple accounts logged in or are part of an org, you can specify which acc/org you want the runs to go to.
  • btw, login info is stored in ~/.netrc.

Things to log

During training loop:

• Loss:
• Graphs?

Misc.

Artifacts

“Conceptually, an artifact is simply a directory in which you can store whatever you want, be it images, HTML, code, audio, or raw binary data.”

Good for storing:

• Saving plots of test images
• Datasets and pickled files
• Model object files

For dataset and model saving, use artifacts.

Full code changes

In training loop:

I am confused about

• What wandb.config provides