calculatedcontent/weightwatcher is a Python diagnostic tool that analyses neural networks without access to training or test data. Its purpose is to predict a model's test accuracy and find signs of overfitting from weight statistics alone, and the most recent release is 0.7.3.5.
It analyses the Conv2D and Dense layers of pretrained or still-training pyTorch and Keras models, reads their weight spectra and compares test accuracy across models. It can look for signs of overfitting or excessive parameterization, and it also detects the problems that may arise during model compression or fine-tuning, together with signs of over- and under-fitting. The README states that this method is based on the Theory of Heavy-Tailed Self-Regularization (HT-SR). The license is Apache-2.0, the repository was created on 28 November 2018, and the star count is 1778.
That said, the given material does not lay out the changes in the 0.7.3.5 release. The accuracy predictions are estimates that follow the method described in the README, not values measured against test data, and the layers the README names as analysis targets are limited to Conv2D and Dense. How well the predictions hold for models of other structures, and how far they diverge from real performance, is not stated in the material. Judged carefully, it is closer to a tool for sweeping over the state of a model without touching training data.