Stable Invariant Models via Koopman Spectra

This repository contains source code for Stable Invariant Models via Koopman Spectra.

Dependencies

We tested the implementation with the following versions.

• Python 3.8.8
• Numpy 1.19.5
• Pytorch 1.8.1
• Ray 1.3.0
• Imageio 2.9.0

Usage

The code is located in src. For example, run_imgcl.py runs the single-tier stable invariant model in the image classification task:

python3 run_imgcl.py --sfix --tune --dataset mnist sim-single --basis cnn

• --sfix option sets a specific seed to control randomness.
• --tune option runs the code together with hyperparameter tuning via Ray Tune.
• --dataset option specifies an evaluated dataset.
• sim-single is replaced with sim-two if you want to run the two-tier stable invariant model.
• --basis option determines the basis function (for now, the only convolutional network (cnn) is implemented for the image classification task).

If you want to reduce the size of log files for Ray Tune, set TUNE_MAX_PENDING_TRIALS_PG to 1. Ray Tune also uses multiple GPUs by default. Setting CUDA_VISIBLE_DEVICES can reduce available GPUs. For example,

TUNE_MAX_PENDING_TRIALS_PG=1 CUDA_VISIBLE_DEVICES=0 python3 run_imgcl.py --sfix --tune --dataset mnist sim-single --basis cnn

Please see run_imgcl.py for more details. The following scripts in src describe running examples involving the other two tasks in the paper:

• cpmem.sh: copy memory task
• imgcl.sh: image classification task
• imgreg.sh: image regression task