Question regarding the iteration number

[ericlearning]

It seems like in line 80 of train.py, the step is changed only when used_sample > phase x 2. This makes sense in resolution >= 16x16, because each step is transition + stablization.

But in 8x8, alpha=1 all the time. So if I am understanding correctly, all transitions and stabilization are trained for "phase" (600,000) samples, except the 8x8 stabilization for phase x 2 (1,200,000) samples, right?

Is this behavior consistent with the original implementation? I expected 8x8 stablization to also train for "phase" samples.

Also, thank you for this amazing repo!

[rosinality]

Somewhat different. In the official implementations initial resolutions trained without phase transition steps, and trained with 600k images.