Skip to content

Conversation

@virajbshah
Copy link
Contributor

@virajbshah virajbshah commented Apr 14, 2025

  • Adds a call to a new no-op method to set up histogram summaries to the model initialization pipeline.
  • Overrides this in GnnModelBase to log histogram summaries for all of its trainable variables, i.e. those in the graph network layers.

@virajbshah virajbshah requested a review from ondrasej April 14, 2025 20:39
@virajbshah virajbshah force-pushed the histogram-summaries branch from 5a46a8c to 572fa73 Compare June 23, 2025 11:30
@virajbshah virajbshah requested a review from ondrasej June 23, 2025 14:05
Copy link
Collaborator

@boomanaiden154 boomanaiden154 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What exactly is the purpose of logging the weight distributions like this?

@virajbshah
Copy link
Contributor Author

It's supposed to be an extra diagnostic to catch things like some weights exploding or gradients vanishing. I added it to compare embedding distributions between regular and context nodes.

@boomanaiden154
Copy link
Collaborator

It's supposed to be an extra diagnostic to catch things like some weights exploding or gradients vanishing. I added it to compare embedding distributions between regular and context nodes.

Ah, okay. Definitely makes sense as a diagnostic tool for that. And I guess the size increase isn't large because it's just a histogram over each layer.

Copy link
Collaborator

@boomanaiden154 boomanaiden154 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, but it's probably good to wait for Ondrej's approval as well.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants