search

kfallah / NODE-Denoiser

Experiments on Neural ODEs for Denoising and Other Problems
4 stars 2 forks source link

Performance #1

Open xiaoiker opened 3 years ago

xiaoiker commented 3 years ago

Hey, Thanks for your codes. Can I ask about the performance of your ODE-denoiser? Compared to a commmonly used model like Unet for image desnoising, is there any advangteges in terms of both computational speed and performance?

Thanks.

kfallah commented 3 years ago

Hi, It was not able to outperform a Unet and took a significant amount of time to train. That being said, I did not invest a lot of time into architecture search for different ODE functions.

My intuition is that an ODE-denoiser would be most effective with irregularly-spaced measurements that need to be denoised (i.e. MRI measurements over time that are taken at irregular time steps). If you have regular time steps, it is not clear to me that you would benefit over using an RNN.

Best, Kion