ebanflo42/rnn-kernel

Framing RNN as a kernel method: A neural ODE approach

Requirements

For reproductibility purposes, we advise to install the project in a dedicated virtual environment to make sure the specific requirements are satisfied. Recommended Python version: 3.8.x.

To install requirements:

pip install -r requirements.txt

Additional pip requirements:

pip install tqdm joblib librosa prefetch_generator python_speech_features

In addition, this package requires installing signatory, which is not entirely straightforward. The following command should work on Linux and Windows:

pip install signatory==1.2.4.1.7.1 --no-cache-dir --force-reinstall

Note that the default MacOS C++ compiler does not support openmp, which is required to compile signatory. A solution is to install the llvm compiler, then set the environment variables so that pip uses this compiler. The following snippet should work.

brew install llvm libomp
export CXX=/usr/local/opt/llvm/bin/clang
export LD=/usr/local/opt/llvm/bin/clang
export CC=/usr/local/opt/llvm/bin/clang
pip install signatory==1.2.4.1.7.1 --no-cache-dir --force-reinstall

For more detail, see their documentation.

Run google speech commands experiment

python main.py --adversarial-exp google_speech

Download speech commands at: http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz

Set the gsc_path in adversarial_experiment.py to /path/to/speech_commands_v0.02.

Testing

python main.py --taylor-exp test
python main.py --adversarial-exp test

Reproducing the paper figures

python main.py --taylor-exp final
python main.py --adversarial-exp spirals_penalization