michedev / ddpms-pytorch Goto Github PK

should not it be: vlb += torch.distributions.kl_divergence(q, p) * t_eq_last.float()

Hello, thank you for the code. This code seems to work fine when there is a square image with H=W but in cases where H and W are different, the network confuses H and W channel in the upsampling step of unet (ddpm_pytorch/model/unet_class.py) and throws an error:

File "/home/DDPMs-Pytorch/ddpm_pytorch/model/unet_class.py", line 119, in forward_unet
h = torch.cat([h, hs[-i]], dim=1)
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 62 but got size 72 for tensor number 1 in the list.

62 is Width and 72 is the height of downsampled image with orig size of 288 x 248. Please let me know if my observation is correct or I made a mistake somewhere. Thanks.