DS-DDPM

Domain Specific Denoising Diffusion Probabilistic Models for Brain Dynamics/EEG Signals

---

A minimum implementation of Domain Specific Denoising Diffusion Probabilistic Models for Brain Dynamics/EEG Signals

The intuition is simple, traditional EEG denoising methods normally apply blind source separation or frequency filtering. We apply domain generation to model human subject's difference separation as a summation of domain generation and clean content denoising as the figure shows below.

To better illustrate the generated brain activity and compare it with the real sample, give an illustration below. The first row is the EEG sensor spatial relations (left for 3D spatial position, right for 2d projection). The second raw is the signal variance through time ( left for generated, right for real data).

Citation

@article{duan2023domain,
  title={Domain-specific denoising diffusion probabilistic models for brain dynamics},
  author={Duan, Yiqun and Zhou, Jinzhao and Wang, Zhen and Chang, Yu-Cheng and Wang, Yu-Kai and Lin, Chin-Teng},
  journal={arXiv preprint arXiv:2305.04200},
  year={2023}
}

---

Installation

Please install required pip packages

pip install tensorboardx
pip install labml
pip install scikit-learn
pip install torch torchvision
pip instlal -r requirements.txt

If warning of any unfound versions , just install the latest version with conda.

Training

The training entry scripts:

python unet2d_overlap.py

If you want use wandb to log metrics on webiste first run init

wandb login

Please note that the currrent setting stacks time segments of EEG signals to form a long time series. The default setting is 244 time window with stride step 75. You can change it in the training file.

One important problem is multiple logits for each time window may lead to multiple values for a same time point. We use a constraint to constraint different values associated to a same time point to be the same. You can use time_diff_constraint=True to control the constraint. To give a illustrate, we give a comparision example of the generated EEG signals with and without the constraint, where red denotes training with time diff constraint, green denotes training without time diff constraint. It illustrates the efficiency of our loss design.

Inference

Please config args as below as you training file give. Also set the runid to your trained checkpoints folder. Given the example below please set runid=34854f3ed38711edb808e4434b7714aa A sample structure under the log/experiment_name folder could be:

.
└── 34854f3ed38711edb808e4434b7714aa
    ├── checkpoints
    ├── configs.yaml
    ├── indicators.yaml
    ├── pids
    ├── run.yaml
    ├── source.diff
    └── wandb

The inference entry scripts:

python sample_save.py

The generated example of denoised EEG signals (blue) with separated noise (green). This method may also be used for other time sequence signals. Below is a sampled animation of the generated process of sampling sythetic EEG signals from noise given random noise and subjects. The noise decrease through time steps. The clean signal then rapidly take the lead of the whole process.