Intermediate Activations – the forward hook (2020)(web.stanford.edu)

web.stanford.edu

Intermediate Activations – the forward hook (2020)

https://web.stanford.edu/~nanbhas/blog/forward-hooks-pytorch/

4 comments

knlb20222년 전

I gave a small talk on how to really push using hooks for logging intermediate values (including capturing gradients from torch & fx scripted modules) that may be useful: https://static.sched.com/hosted_files/pytorch2023/40/Interme...

jph002년 전

FYI, for anyone interesting in creating and using hooks to better understand what's happening in your model, I created a free lesson covering that:

https://course.fast.ai/Lessons/lesson17.html

jey2년 전

(2020)

vinay4272년 전

For those interested in playing with or doing research using model internals, the TransformerLens [1] project appears to be the leading open-source tooling in this area. It allows for loading dozens of different models, adding hooks, displaying activations in a format compatible with CircuitsVis, and other (mechanistic) interpretability work.

[1] https://github.com/neelnanda-io/TransformerLens

[2] https://github.com/alan-cooney/CircuitsVis