Open dmarx opened 2 years ago
re: noise annealing (via https://github.com/LAION-AI/notebooks/blob/main/DALLE2-Prior%2BDeep-Image-Prior.ipynb):
noise_ramp = 1 - min(1, itt / iterations)
net_input_noised = net_input
if input_noise_strength:
phi = min(1, noise_ramp * input_noise_strength) * math.pi / 2
noise = torch.randn_like(net_input)
net_input_noised = net_input * math.cos(phi) + noise * math.sin(phi)
with torch.cuda.amp.autocast():
out = net(net_input_noised * input_scale).float()
current blocker: a lot of dependent code assumes latent is a single tensor. downstream operations on the image representation attempt to call methods on it like "clone" that don't have good analogs with the module dict/list.
I think the solution is instead of adding special cases, generalize the other image reps to expect containers with arbitrarily many image representation components of heterogeneous sizes.
updating EMAImage to operate on parameter dicts/lists could potentially solve a good chunk of this
migrate to branch dip_ema_simple
still to do: