wildphoton / randconv

Code for ICLR2021 paper "Robust and Generalizable Visual Representation Learning via Random Convolutions"

RandConv

Official repo for Robust and Generalizable Visual Representation Learning via Random Convolutions (ICLR2021)

Update 05/10: Code for RandConv and training scripts on digits data are available now! Scripts for PACS and imagenet are on the way.

Requirements

See requirements.txt. Note that Pytorch v1.7 was used for testing.

Running RandConv on Digits data

• MNIST-C has to be manually downloaded from https://github.com/google-research/mnist-c. Unzip the data into ./data/MNIST-M or change the data path in train_digits.py.
• exp_mnist10k.sh provided bash commands for reproduce digits experiments in the paper. You can select the specific settings by (un)commenting lines. bash exp_mnist10k.sh 0 will run selected settings on GPU 0.