Code for Generating realistic neurophysiological time series with denoising diffusion probabilistic models.
The repository contains all the scripts to run the experiments presented in the paper.
To run the experiments, first download the datasets (links in the paper) and change the filepath parameter in the dataset configuration files in conf/dataset. To save all outputs locally, set the save_path parameter in conf/base to the desired location.
Experiment tracking and saving with Weights & Biases is also supported, but disabled by default.
Finally, use the shell scripts provided in scripts to run the experiments!
Additionally, an example jupyter notebook example_ner_notebook.ipynb, where a diffusion model is trained to generate the BCI Challenge @ NER 2015 data, is provided together with the data.
Install dependencies via pip:
git clone https://github.com/mackelab/neural_timeseries_diffusion.git
cd neural_timeseries_diffusion
pip install -e .This requires writing a dataloader for your dataset. Example dataloaders (for all the datasets used in paper) can be found in
datasets.py.
Two denoiser different denoiser architectures based on structrured convolutions are provided, which only differ in the way conditional information is passed to the network.
In general, the architecture using adaptive layer norm (AdaConv) is preferable and a good first choice over the one where conitional information is just concatenated (CatConv) (see this paper for a comparison of both approaches for an image diffusion model).
The standard white noise is a good initial choice. However, using the OU process can improve the quality of the generated power spectra. This requires setting the length scale of the OU process as an additional hyperparameter in the diffusion_kernel config files.
After downloading the datasets and updating the file paths, run the experiment shell scripts:
cd scripts
./<name_of_script>.shThis produces result files, which can then be passed to the plotting notebooks.