search

AI4HealthUOL / SSSD-ECG

Repository for the paper: 'Diffusion-based Conditional ECG Generation with Structured State Space Models'
MIT License
42 stars 6 forks source link

DIffusion hyperparameters #15

Closed gabrielvc closed 6 months ago

gabrielvc commented 6 months ago

Hi,

As far as i'm concerned, there is an issue with the current diffusion hyperparameters defined in the config file. Running the current parameters for T=200 yield an sqrt(alpha_bar) of approximately 0.4, which is of course far from the Gaussian, which would be equivalent to sqrt(alpha_bar) of approximately 0.01 or something really small. I guess this is why I can't generate ECGs using the generate function, even if my trained loss is extremely small. Indeed, the network is able to denoise moderately noisy ecgs but generation still don't work.

Furthermore, I could not find in your paper the hyperparameters you used in the diffusion training. I would really like to be able to reproduce your work, but for the moment it is really challenging.

juanlopezcode commented 6 months ago

Hi, The network should run with the provided diffusion hyperparameters, we mentioned them in the appendix of our paper, but they are also under the config file in this repo: https://github.com/AI4HealthUOL/SSSD-ECG/blob/main/src/sssd/config/config_SSSD_ECG.json

Follow the train.py and inference.py files under the following directory, these should be your unique files to run once dataset is located (don't update diffusion hyperparameters): SSSD-ECG/src /sssd/

I hope this helps Best Juan