about the image size

[Yuchen-Hou0210]

i choose ffhq-dataset's thumbnails128x128. I'm curious about which parameters need to be modified？should the --max_size in train.py be modified? should the number of layers in model.py be modified, such as only 6 layers instead of 9 layers?

[rosinality]

You only need to adjust --max_size arguments. Number of layers will be adjusted accordingly.

[Yuchen-Hou0210]

You only need to adjust --max_size arguments. Number of layers will be adjusted accordingly.

thank you very much！And I I have another question：what is the function of FusedDownsample and Fusedupsample ? Why we need it if the resolution is more than 64px, but not if the resolution is less than 64px? I'm a little confused。

[rosinality]

FusedUp(Down)sample is equivalent to conv + avgpool or nearest neighbor upsample + conv. In the official implementations it is used if resolution is larger than 64px as it is more performant at that resolutions. But I haven't measured the performance differences, though.

[Yuchen-Hou0210]

FusedUp(Down)sample is equivalent to conv + avgpool or nearest neighbor upsample + conv. In the official implementations it is used if resolution is larger than 64px as it is more performant at that resolutions. But I haven't measured the performance differences, though.

ok! Thanks, it helps me a lot