bayesian-sam

Code for SAM as an Optimal Relaxation of Bayes, T. Moellenhoff, M. E. Khan, ICLR 2023.

installation and requirements

The code requires JAX and various other standard dependencies such as matplotlib and numpy; see the 'requirements.txt'.

To train on TinyImageNet, you will need to download the dataset from here and extract it into the datasetfolder directory (see the 'data.py' file).

If you have wget installed, you can download the dataset via

wget http://cs231n.stanford.edu/tiny-imagenet-200.zip

examples

ResNet-18 on CIFAR-10

SGD:

python3 train.py --alpha 0.03 --beta1 0.9 --priorprec 25 --optim sgd --dataset cifar10

This should train to around ~94.8% test-accuracy.

SAM-SGD:

python3 train.py --alpha 0.03 --beta1 0.9 --priorprec 25 --rho 0.05 --batchsplit 8 --optim sam --dataset cifar10

This should train to around ~95.7% test-accuracy.

bSAM:

python3 train.py --alpha 0.5 --beta1 0.9 --beta2 0.999 --priorprec 40 --rho 0.01 --batchsplit 8 --optim bsam --dataset cifar10 --dafactor 4

This should train to around ~96% test-accuracy.

To test our trained models and evaluate nll, ECE and AUROC we can use the following commands:

python3 test.py --resultsfolder results/cifar10_resnet18/sgd/run_0 
python3 test.py --resultsfolder results/cifar10_resnet18/sam/run_0 
python3 test.py --resultsfolder results/cifar10_resnet18/bsam/run_0 --testmc 32

This should produce around the following results

SGD: testacc=94.77%, nll=0.2058, ece=0.0306, auroc=0.9336
SAM: testacc=95.69%, nll=0.1431, ece=0.0134, auroc=0.9468
bSAM: testacc=95.95%, nll=0.1239, ece=0.0082, auroc=0.9536

ResNet-18 on CIFAR-100

Similary, we can train on other datasets; here is for example bSAM on CIFAR-100:

python3 train.py --alpha 0.5 --beta1 0.9 --beta2 0.999 --priorprec 40 --rho 0.01 --batchsplit 8 --optim bsam --dataset cifar100 --dafactor 4

The run should converge to around ~80.4% test accuracy.

ResNet-20 on TinyImageNet

To train on TinyImageNet, download the dataset using the link given above. Then extract the zip-File into the datasets directory and run

python3 setup_tinyimagenet.py

Now, to train SGD on TinyImageNet we use:

python3 train.py --alpha 0.1 --beta1 0.95 --priorprec 20 --optim sgd --dataset tinyimagenet --model resnet20

This should train to around ~52.2% test-accuracy.

SAM-SGD:

python3 train.py --alpha 0.1 --beta1 0.95 --priorprec 20 --rho 0.01 --batchsplit 8 --optim sam --dataset tinyimagenet --model resnet20

This should train to around ~52.4% test-accuracy.

bSAM:

python3 train.py --alpha 0.25 --beta1 0.9 --beta2 0.999 --priorprec 25 --rho 5e-5 --batchsplit 8 --optim bsam --dataset tinyimagenet --model resnet20

This should train to around ~53.5% test-accuracy.

To test the trained models, we can run:

python3 test.py --resultsfolder results/tinyimagenet_resnet20/sgd/run_0 
python3 test.py --resultsfolder results/tinyimagenet_resnet20/sam/run_0 
python3 test.py --resultsfolder results/tinyimagenet_resnet20/bsam/run_0 --testmc 32

This should produce around the following results

SGD: testacc=52.28%, nll=2.0077, ece=0.0452, auroc=0.8316
SAM: testacc=52.39%, nll=1.9961, ece=0.0317, auroc=0.8304
bSAM: testacc=53.37%, nll=1.9594, ece=0.0206, auroc=0.8272

illustrative 2D example

To illustrate bSAM on the above simple 2D example, run the code

python3 demo2d.py

This script saves a file animation.gif, which is the animiation shown above.

troubleshooting

Please contact Thomas if there are issues or quesitons about the code, or raise an issue here in this github repository.

team-approx-bayes/bayesian-sam