cientgu/GIQA

Performing GMM is time-consuming

zwx8981 opened this issue · 2 comments

zwx8981 commented

Hi, it seems that performing GMM on all 70000 images of FFHQ is quite time-consuming. Could you tell me how long did you get the resulting GMM model on FFHQ ?

cientgu commented

Good question. It's really time-consuming because the EM is computed on the CPU instead of the GPU. To me, it cost about 1-1.5 day. The time depends on 1) what CPU you use, 2) the number of kernels in your GMM model 3) the feature dimension.
You may apply a PCA rate to decrease the channel from 2048 to about 700. Also, you may directly apply this GMM70-FFHQ model (https://drive.google.com/file/d/1TjEPTNKpai236cFOotiwYE7cVRuGNzng/view?usp=sharing).
Hope it may help you.

  • 👍1
zwx8981 commented

Thanks for your quick reply!