just building.

github

Akshath Mangudi  • Peerlist

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: <ul><li>detecting corruption (partial failures, etc) </li><li>extracting per-layer metrics (mean, std, etc) </li><li>get distribution stats for very large checkpoint and not crash your computer while doing it. </li><li>diagnostics for unhealthy layers within the checkpoint. </li></ul>To try: <ul><li>setup a virtual environment and run `pip install weightlens`. </li><li>try Weightlens by using `lens analyze &lt;filename&gt;.pth`</li></ul>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: <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="https://peerlist.io/scroll/post/ACTHMQ6R7K6BA8ENRFBPQPRBPR8L98">https://peerlist.io/scroll/post/ACTHMQ6R7K6BA8ENRFBPQPRBPR8L98</a>

Weightlens

[UPDATE]: We have launched our demo for people to try it out! Deep learning workflows today are still fragmented. Training lives in one place. Evaluation scripts live somewhere else. Explainability is often bolted on last. Refrakt is a unified platform that brings: • Model training • Evaluation • Explainability into a single system. Platform access is coming soon! We’ve put together a short 3-minute walkthrough of the UI: <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="https://lnkd.in/gcavFQ3y">https://lnkd.in/gcavFQ3y</a> Join the waitlist: <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="https://refrakt.akshath.tech/">https://refrakt.akshath.tech/</a>