What's the training loss (as opposed to validation loss) like?

Specifically, I’m interesting in getting a loss curve (loss vs. training iteration) for the 200k CGCNN S2EF split. Going to try training and opening with tensorboard, but it’s my first time using tensorboard. Didn’t see much by way of training loss in the papers (maybe I missed it).

Both the training loss and validation loss use the following functional form:

Tensorboard and wandb will by default log all the metrics, allowing you to easily visualize the loss curves. (wandb is a lot more user friendly if you want to try that, simply modify the logger line to wandb in your specified config).

Let me know if that answers your question.

Thank you, this is great. I was able to get tensorboard working (turned out whatever version of Mozilla Firefox wasn’t playing nicely on the compute cluster I was using, switching to Google Chrome fixed the issue).

1 Like