Provide the model type--config-name to train and test models configured as those shown in the paper.

1. This repository depends on a few submodules. To clone either use git clone --recurse-submodules or use git submodule update --init --recursive after cloning.
2. Install the requirements using conda by running conda env create -f environment.yaml

Scripts assume the project root is the current directory.

Example commands:

python sl_pipeline.py --config-name classical +dataset=MNIST

Tensorboards are saved to run_data/tensorboards and can be viewed by running:

tensorboard --logdir ./run_data/tensorboards --reload_multifile True

Only the model with the best validation error is saved. To quickly verify the the test error of this model, run the adversarial robustness script. It prints the nominal test error before performing the attack.

Assuming the current directory is robustness. Notice that the model file name will be different depending on the dataset and model combination you have run. The path provided should provide an idea of the directory structure where models are stored.

These scripts will print the testing error, followed by the testing error with and adversarial attack. Notice adversarial testing requires significantly more resources.

PYTHONPATH=../ python untargeted_robustness.py --config-name classical norm="2" \
+dataset=MNIST \
"+model_file='../run_data/tensorboards/d.MNIST_m.ClassicalModule(RESNET18)_b.128_lr.0.01_wd.0.0001_mepoch120._sd0/default/version_0/checkpoints/epoch=7-step=3375.ckpt'"

PYTHONPATH=../ python untargeted_robustness.py --config-name classical \
norm="inf"  +dataset=MNIST \
"+model_file='../run_data/tensorboards/d.MNIST_m.ClassicalModule(RESNET18)_b.128_lr.0.01_wd.0.0001_mepoch120._sd0/default/version_0/checkpoints/epoch=7-step=3375.ckpt'"

• MNIST
• FashionMNIST
• CIFAR10
• CIFAR100

• anode : Data-controlled dynamics with ResNet18 Component trained through solution differentiation
• classical: ResNet18
• lyapunov: Data-controlled dynamics with ResNet18 Component trained with LyaNet
• continuous_net: ContinuousNet from [1] trained through solution differentiation
• continuous_net_lyapunov: ContinuousNet from [1] trained with LyaNet

1. Continuous-in-Depth Neural Networks Code
2. Learning by Turning: Neural Architecture Aware Optimisation Code
3. LyaNet: A Lyapunov framework for training neural ODEs