"Predicting trends in the quality of state-of-the-art neural networks without access to training or testing data"
Python 3.7.6 weightwatcher 0.2.7 (or ww 0.4 with ww2x, and min_size = 50)
Conda environment in requirements.txt
Jupyter Notebooks for reproducing most Tables and all Figures
All results can be generated using pretrained models available in the torchvision pyTorch models (except ResNet-1K, which requies the Cv Sandbox)
Contains data from weightwatcher runs using Google Colab All Tables and Figures are generated directly from this raw data
Jupyter Notebooks for reproducing Figure 4 and accompanying text (note: user must install Intell distiller to run these)
original Latex files
images, generated by Jupyter Notebooks
current PDF of the archive paper
The original weightwatcher calculations were done in the Summer of 2019, and then repeated in Jan 2020 using more pretrained models (from the OSMR repo)
Since that time, the weightwatcher code has been updated, and the OSMR models have have changed
This paper reports details results from the Jan 2020 data, stored in data/omsr
The calculations can be repeated using weightwatcher (with ww2x=True set) however, there may be minor differences in the numerical results.
Data from older submission: Google Colab Notebooks for reproducing results in sections 6
Notebooks can be run in parallel on the users Google Cloub They Will download pretrained models from the CV Sandbox