michaelsdr/jacobian_free_backprop

Implicit networks can be trained efficiently and simply by using Jacobian-free Backprop (JFB).

Jacobian-Free Backprop

Associated Publication

JFB: Jacobian-Free Backpropagation for Implicit Models (arXiv Link)

@article{WuFung2022JFB,
    title={JFB: Jacobian-Free Backpropagation for Implicit Models},
    author={Fung, Samy Wu and Heaton, Howard and Li, Qiuwei and McKenzie, Daniel and Osher, Stanley and Yin, Wotao},
    journal={Proceedings of the AAAI Conference on Artificial Intelligence},
    year={2022}
}

Set-up

Install all the requirements:

pip install -r requirements.txt 

Training

For each dataset, there are three types of training drivers:

1. FPN with our proposed backprop:

	python train_CIFAR10.py
	python train_CIFAR10_Unaugmented.py
	python train_MNIST.py
	python train_SVHN.py

2. FPN with Jacobian-based backprop:

	python train_CIFAR10_Jacobian_Based.py
	python train_CIFAR10_Unaugmented_Jacobian_Based.py
	python train_MNIST_Jacobian_Based.py
	python train_SVHN_Jacobian_Based.py

3. Explicit models.

	python train_CIFAR10_Explicit.py
	python train_CIFAR10_Unaugmented_Explicit.py
	python train_MNIST_Explicit.py
	python train_SVHN_Explicit.py