Open chaibosong opened 3 months ago
Sorry for the confusion. We have corrected the num_images
parameter in the revised README.md
; it should be set to 50,000 instead of 500,000. Testing the FID results took us approximately 4-5 hours using 8 A100 40GB GPUs with a total batch size of 2000. It is advisable to utilize as much memory as possible on each GPU. Additionally, the number of images used for evaluating the FID results is configurable; you can set this number according to your needs, e.g., 10,000 images. However, please note that using too few images may affect the accuracy of the FID scores.
Thank you very much for your prompt reply!Another small question, the cifa weight file you provided is about 1.9G, and the weight I trained myself is 763M. What is the reason for this?
Please see issue #6.
we use "CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python main_SDDPM.py --eval --pre_trained_path './SDDPM_CIFAR10.pt' --num_images 500000 --fid_cache './stats/cifar10.train.npz' " with 7 A100 GPUs,
and the eval time is
generating images: 1%|█ | 30/3907 [1:11:32<152:46:18, 141.86s/it]
I think there must be some problem.