Inference sampling procedure

[andreabosisio]

BrepGen/sample.py

Lines 128 to 129 in ee8f342

	pndm_scheduler.set_timesteps(200)
	for t in tqdm(pndm_scheduler.timesteps[:158]):#

BrepGen/sample.py

Line 137 in ee8f342

surfPos = pndm_scheduler.step(pred, t, surfPos).prev_sample

I can't understand why here the timesteps are cut to 158, and, most importantly, why the output of this PNDM denoising process is then fed to another DDPM process starting from the lasts 250 steps:

BrepGen/sample.py

Lines 144 to 149 in ee8f342

	ddpm_scheduler.set_timesteps(1000)
	for t in tqdm(ddpm_scheduler.timesteps[-250:]):
	timesteps = t.reshape(-1).cuda()
	if class_label is not None:
	_surfPos_ = surfPos.repeat(2,1,1)
	pred = surfPos_model(_surfPos_, timesteps, class_label)

Could you please explain the motivation and intended effects of these two aspects of the overall process?