Weightlens is an analysis tool that analyzes your checkpoints before working with them.

If you have worked with PyTorch or model checkpoints in general, you will understand just how frustrating it is to work with checkpoints that may or may not be corrupted, losing hours of time.

Weightlens fixes this by:

• detecting corruption (partial failures, etc)

• extracting per-layer metrics (mean, std, etc)

• get distribution stats for very large checkpoint and not crash your computer while doing it.

• diagnostics for unhealthy layers within the checkpoint.

To try:

• setup a virtual environment and run `pip install weightlens`.

• try Weightlens by using `lens analyze <filename>.pth`

NOTE: This is under active development, and this library is not perfect. But any feedback will help me make Weightlens a much better library than what it is.

Check out what I solved yesterday: https://peerlist.io/scroll/post/ACTHMQ6R7K6BA8ENRFBPQPRBPR8L98