search

mirthAI / Fast-DDPM

MIT License
52 stars 7 forks source link

How to change the training process? #2

Closed Lrihao0413 closed 1 month ago

Lrihao0413 commented 2 months ago

First of all, congratulations on the great work, and thanks for making it open-source!

I'm in doubt about how to accelerate the training process? I have tried adjusting some parameters, but only accelerated the sampling speed, but the training speed remains the same. I hope to receive your answer. Thank you!

Sebastianjhx commented 2 months ago

Hi there, thanks for your interest in our works. Fast-DDPM speeds up the training process by achieving the better performance more quickly, in other words faster convergence. For more details, please refer to the result section in our paper. For instance in the task of image denoising, Fast-DDPM trained for 200,000 iterations could easily reach PSNR 37.0 and SSIM 0.91, while DDPM only reaches the best results of 35.4 and 0.87 at the iteration 600,000.