Paper "Music Generation and Transformation with Moment Matching-Scattering Inverse Networks", M. Andreux, S. Mallat, ISMIR 2018
- Install the conda environment:
conda env create -f conda-env.yml - Install the local package:
pip install -e .
Go to scattering_autoencoder/datasets/toy_model_gen.py, and call it with a --path argument specifying the place where the dataset will be stored and --n_examples, the number of examples in each train/test set.
- Download some existing dataset here: train/test
- Compute their scattering with
experiments/music/preprocess_dataset.py, with arguments: **--path_data: path to the training data (music_train_norm2_4096.npy) **--path_save: root folder where to save the preprocessed data. ** Note that this step can take up to 6 hours without CPU parallelization. - Renormalize the channels using code from the notebook in
experiments/music/renormalize_datasets.ipynb
The relevant file for training is experiments/main.py. It takes as argument
--argfile, which should be a json file containing parameters for the experiments. Examples thereof are provided in experiments/music and experiments/toy.
- First block: "J" -> "dataset_name": arguments related to the scattering used as input to the convolutional network
- Second block: "loss_type" -> "p_order": arguments related to the loss function.
- Third block: "last_convolution" -> "size_channel_1": arguments related to the architecture of the network
- Fourth block: "lr" -> "gamma": arguments related to the training procedure
- Fifth block: "is_cuda" -> "timestamp_data": arguments related to the dataset. In particular, the argument "timestamp_data" should be updated according to the dataset which is used.