PaRT: Parallel Learning for Robust and Transparent AI

This repository contains the code for PaRT, an algorithm for training a base network on multiple tasks in parallel. The diagram of PaRT is shown in the figure below.

Below, we provide details regarding dependencies and the instructions for running the code for each experiment. We have prepared scripts for each experiment to help the user have a smooth experience.

Dependencies

• python >= 3.8
• pytorch >= 1.7
• scikit-learn
• torchvision
• tensorboard
• matplotlib
• pillow
• psutil
• scipy
• numpy
• tqdm

SETUP ENVIRONMENT

To setup the conda env and create the required directories go to the scripts directory and run the following commands in the terminal:

conda init bash
bash -i setupEnv.sh

Check that the final output of these commands is:

Installed torch version {---}
Virtual environment was made successfully

CIFAR 100 EXPERIMENTS

Instructions to run the code for the CIFAR100 experiments:

--------------------- BASELINE EXPERIMENTS ---------------------

To run the baseline experiments for the first seed, go to the scripts directory and run the following command in the terminal:

bash -i runCIFAR100Baseline.sh ../../scripts/test_case0_cifar100_baseline.json

To run the experiment for other seeds, simply change the value of test_case in test_case0_cifar100_baseline.json to 1,2,3, or 4.

--------------------- PARALLEL EXPERIMENTS ---------------------

To run the parallel experiments for the first seed, go to the scripts directory and run the following command in the terminal:

bash -i runCIFAR100Parallel.sh ../../scripts/test_case0_cifar100_parallel.json

To run the experiment for other seeds, simply change the value of test_case in test_case0_cifar100_parallel.json to 1,2,3, or 4.

CIFAR 10 AND CIFAR 100 EXPERIMENTS

Instructions to run the code for the CIFAR10 and CIFAR100 experiments:

--------------------- BASELINE EXPERIMENTS ---------------------

To run the parallel experiments for the first seed, go to the scripts directory and run the following command in the terminal:

bash -i runCIFAR10_100Baseline.sh ../../scripts/test_case0_cifar10_100_baseline.json

To run the experiment for other seeds, simply change the value of test_case in test_case0_cifar10_100_baseline.json to 1,2,3, or 4.

--------------------- PARALLEL EXPERIMENTS ---------------------

To run the baseline experiments for the first seed, go to the scripts directory and run the following command in the terminal:

bash -i runCIFAR10_100Parallel.sh ../../scripts/test_case0_cifar10_100_parallel.json

To run the experiment for other seeds, simply change the value of test_case in test_case0_cifar10_100_parallel.json to 1,2,3, or 4.

FIVETASKS EXPERIMENTS

The dataset for this experiment can be downloaded from the link provided by the CPG GitHub Page or Here. Instructions to run the code for the FiveTasks experiments:

--------------------- BASELINE EXPERIMENTS ---------------------

To run the baseline experiments for the first seed, go to the scripts directory and run the following command in the terminal:

bash -i run5TasksBaseline.sh ../../scripts/test_case0_5tasks_baseline.json

To run the experiment for other seeds, simply change the value of test_case in test_case0_5tasks_baseline.json to 1,2,3, or 4.

--------------------- PARALLEL EXPERIMENTS ---------------------

To run the parallel experiments for the first seed, go to the scripts directory and run the following command in the terminal:

bash -i run5TasksParallel.sh ../../scripts/test_case0_5tasks_parallel.json

To run the experiment for other seeds, simply change the value of test_case in test_case0_5tasks_parallel.json to 1,2,3, or 4.

Paper

Please cite our paper:

Paknezhad, M., Rengarajan, H., Yuan, C., Suresh, S., Gupta, M., Ramasamy, S., Lee H. K., PaRT: Parallel Learning Towards Robust and Transparent AI, arXiv:2201.09534 (2022)