Major refactorings, plotting notebooks, license change

Author: jsvetter

LICENSE

@@ -1,33 +1,41 @@
 # Generating realistic neurophysiological time series with denoising diffusion probabilistic models
 
-This repository contains research code for the preprint [Generating realistic neurophysiological time series with denoising diffusion probabilistic models](https://www.biorxiv.org/content/10.1101/2023.08.23.554148v1).
+Code for [Generating realistic neurophysiological time series with denoising diffusion probabilistic models](https://www.biorxiv.org/content/10.1101/2023.08.23.554148v1).
 
-It 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](https://wandb.ai/site) is also supported, but disabled by default.  
+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](https://wandb.ai/site) 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.
 
-Coming soon: Notebooks for plotting the figures!
-
-
-## Usage
-
-Install dependencies:
+## Installation
+Install dependencies via pip:
 ```shell
 git clone https://github.com/mackelab/neural_timeseries_diffusion.git
 cd neural_timeseries_diffusion
 pip install -e .
 ```
 
-Run shell scripts:
+## Further information
+
+#### How can apply the DDPM to my datasets?
+This requires writing a dataloader for your dataset. Example dataloaders (for all the datasets used in paper) can be found in 
+`datasets.py`.
+
+#### What options do I have for the denoiser architectures?
+Two denoiser different denoiser architectures based on [structrured convolutions](https://arxiv.org/abs/2210.09298) 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](https://arxiv.org/abs/2212.09748) for a comparison of both approaches for an image diffusion model).
+
+#### White noise vs. OU process
+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.
+
+## Running the experiments
+After downloading the datasets and updating the file paths, run the experiment shell scripts:
 ```shell
 cd scripts
 ./<name_of_script>.sh
 ```
-
-Run the example jupyter notebook:
-- Make sure that the `jupyter` package is installed
-- Open and execute the notebook
-
-
+This produces result files, which can then be passed to the plotting notebooks.

README.md

@@ -6,7 +6,7 @@ run_experiment: True
 use_cuda_if_available: True
 
 wandb_mode: disabled
-wandb_project: time_series_diffusion
+wandb_project: TODO
 wandb_entity: TODO
 home_path: "TODO" # wandb home path
-save_path: null # Local saving
+save_path: null # local saving

conf/base/wandb_default.yaml

@@ -1,7 +1,7 @@
 defaults:
   - base: wandb_default
   - dataset: ner
-  - network: lconv_ner
+  - network: ada_conv_ner
   - diffusion_kernel: white_noise
   - diffusion: diffusion_linear_200
   - optimizer: base_optimizer

conf/config.yaml

@@ -2,7 +2,5 @@ self: ner
 filepath: "../data"
 patient_id: "S02"
 signal_length: 260
-with_time_emb: True
-cond_time_dim: 32
 train_test_split: 1.0
 split_seed: 0

conf/dataset/ner.yaml

@@ -1,4 +1,4 @@
 # @package _global_
 swr_prediction_experiment:
-  channel: 0 # can be integer or list of integers
+  channel: 0
   batch_size: 1000

conf/diffusion/diffusion_cosine_500.yaml

@@ -0,0 +1,13 @@
+self: ada_conv
+signal_channel: 64 # only for P07
+cond_dim: 128 # only for P07
+hidden_channel: 8
+in_kernel_size: 53
+out_kernel_size: 53
+slconv_kernel_size: 53
+num_scales: 4
+num_blocks: 3
+num_off_diag: 8
+padding_mode: circular
+use_fft_conv: False
+use_pos_emb: False

conf/experiments/swr_prediction_experiment/swr_prediction_experiment.yaml

@@ -0,0 +1,13 @@
+self: ada_conv
+signal_channel: 3
+cond_dim: 0
+hidden_channel: 64
+in_kernel_size: 64
+out_kernel_size: 64
+slconv_kernel_size: 64
+num_scales: 5
+num_blocks: 4
+num_off_diag: 64
+padding_mode: circular
+use_fft_conv: False
+use_pos_emb: False

conf/network/ada_conv_ajile.yaml

@@ -0,0 +1,13 @@
+self: ada_conv
+signal_channel: 56
+cond_dim: 0
+hidden_channel: 8
+in_kernel_size: 65
+out_kernel_size: 65
+slconv_kernel_size: 65
+num_scales: 1
+num_blocks: 3
+num_off_diag: 8
+padding_mode: circular
+use_fft_conv: False
+use_pos_emb: True