sam.pytorch

A PyTorch implementation of Sharpness-Aware Minimization for Efficiently Improving Generalization ( Foret+2020) Paper, Official implementation .

Requirements

• Python>=3.8
• PyTorch>=1.7.1

To run the example, you further need

• homura by pip install -U homura-core==2020.12.0
• chika by pip install -U chika

Example

python cifar10.py [--optim.name {sam,sgd}] [--model {renst20, wrn28_2}] [--optim.rho 0.05]

Results: Test Accuracy (CIFAR-10)

Model | SAM | SGD | --- | --- | --- | ResNet-20 | 93.5| 93.2| WRN28-2 | 95.8| 95.4| ResNeXT29 | 96.4| 95.8|

SAM needs double forward passes per each update, thus training with SAM is slower than training with SGD. In case of ResNet-20 training, 80 mins vs 50 mins on my environment. Additional options --use_amp --jit_model may slightly accelerates the training.

Usage

SAMSGD can be used as a drop-in replacement of PyTorch optimizers by using a closure as follows. Also, it is compatible with lr_scheduler and has state_dict and load_state_dict. Currently, this implementation does not support multiple parameter groups.

from sam import SAMSGD

optimizer = SAMSGD(model.parameters(), lr=1e-1, rho=0.05)

for input, target in dataset:
    def closure():
        optimizer.zero_grad()
        output = model(input)
        loss = loss_f(output, target)
        loss.backward()
        return loss


    loss = optimizer.step(closure)

Citation

@ARTICLE{2020arXiv201001412F,
    author = {{Foret}, Pierre and {Kleiner}, Ariel and {Mobahi}, Hossein and {Neyshabur}, Behnam},
    title = "{Sharpness-Aware Minimization for Efficiently Improving Generalization}",
    year = 2020,
    eid = {arXiv:2010.01412},
    eprint = {2010.01412},
}

@software{sampytorch
    author = {Ryuichiro Hataya},
    titile = {sam.pytorch},
    url    = {https://github.com/moskomule/sam.pytorch},
    year   = {2020}
}