There is a discrepancy in the results when reproducing the facades in cyclegan, and the FID and IS are inconsistent with the provided ones

[warning0]

Using 1GPU 1batchsize to reproduce the facades, there is a gap between the image conversion effect and the weight conversion effect provided. And my FID and IS are 119.3 and 1.55.

Also note that dataroot='in config/ data/unpaired_ facades' and domain_ a = 'mask'domain_ B='photo' does not correspond to the dataset downloaded from the document, it should be dataroot='/ data/facades', while swapping the positions of a and b in the folder.

What should I do to improve training effectiveness to achieve consistency with the weights provided

[yftianwin]

Hi, I wanna learn how can I implemente image evaluation metrics, such as FID, SSIM, IS, etc. If you have perfect methods, would you like to share with me? Thanks very much!