Weightlens is here!
If you've worked with models and checkpoints, you will know how frustrating it is to deal with partial downloads, corrupted .pth files, and the list goes on, especially if it's a large project.
To spare the burden for everyone, I have created a small tool that allows you to analyze a model's checkpoints where you can:
- detect corruption (partial failures, tensor access failures, etc)
- extract per-layer metrics (mean, std, l2 norm, etc)
- get global distribution stats which are properly streamed and won't break your computer
- deterministic diagnostics for unhealthy layers.
To try it, run pip install weightlens into your virtual environment and run lens analyze <filename>.pth and check it out!
Please do give it a star if you like it!
I would love your thoughts on testing this out and getting your feedback.