fabiocarrara / neural-ode-features

Pytorch code to train image classifiers based on ODE Nets on MNIST and CIFAR-10, extract features and test robustness to adversarial examples

Neural ODE Image Classifiers

Pytorch code for training and evaluating Neural ODEs image classifiers on MNIST and CIFAR-10 datasets. It reproduces experiments presented in the following papers:

[1] Carrara, F., Amato, G., Falchi, F. and Gennaro, C., 2019, September. Evaluation of Continuous Image Features Learned by ODE Nets. In International Conference on Image Analysis and Processing (ICIAP '19) (pp. 432-442). Springer, Cham.

[2] Carrara, F., Amato, G., Falchi, F. and Gennaro, C., 2020, June. Continuous ODE-defined Image Features for Adaptive Retrieval. In Proceedings of the 2020 International Conference on Multimedia Retrieval (ICMR '20) (pp. 198-206). ACM.

[3] Carrara, F., Caldelli, R., Falchi, F. and Amato, G., 2019, December. On the robustness to adversarial examples of neural ode image classifiers. In 2019 IEEE International Workshop on Information Forensics and Security (WIFS '19) (pp. 1-6). IEEE.

Getting Started

Clone and install requirements:

git clone --recursive https://github.com/fabiocarrara/neural-ode-features.git
cd neural-ode-features
pip install -e torchdiffeq
pip install torchvision foolbox h5py pandas tqdm seaborn sklearn

Reproduce Experiments

To obtain the trained models and reproduce the experiments described in [1] and [2], run

./reproduce.sh

Pre-trained models are also available: neural-ode-features-runs.zip (172MB)

---

To reproduce experiments described in [3], obtain the trained models, and then run

cd adversarial
./reproduce.sh <path/to/specific_run_folder>

to attack a specific model and collect results.

neural-ode-features's Issues

Wondering the best ODE results on CIFAR10 with Odenet of your tests.

Your code works seems really great!
I was confused about the experiments in that paper with MNIST dataset. And the best I can get so far is around 86.5% with ODEnet on CIFAR10. Wondering did you get anything over 90 with ODEnet through your works? I found your code today and it seems you tested way much things than I did.