Missing poses for FFHQ dataset

[Xallt]

Hello,
First of all, thank you for providing all the links to the data you used for your experiments.

After downloading the face poses archive from this link you provide in the README, I noticed that the archive contains 69994 poses, whereas the original FFHQ dataset contains 70000 images.

These are the indices missing from the archive: [43079, 27757, 36783, 51858, 17972, 42709]

Is this intentional, or a bug? You don't have any functionality in your FFHQ dataset implementation to support a mismatch between extracted images and poses.

[YuDeng]

Hi, our face detector failed on these images so we did not use them during training.

[Xallt]

Here when loading the data for the FFHQ dataset you first load all the PNGs and then you load all the corresponding poses for the PNGs. Which means that, if you have missing poses, the loading will inevitably fail.