1D images

[JesseWiers]

Has anyone made the code work for 1D images? I changed in_channels and out_channels to 1 in the creation of the model in script_util.py and removed the conversion to RGB in image_datasets.py. Afterwards, I have been training a model for 80 000 iterations on 14 000 black-white medical images with a batch size of 8 and minibatch of 1 but the samples are pure noise.

[ykcheng9966]

hi，bro！have you solved this problem? Would you mind sharing the parameters for training 1D images？

[leffff]

Please share the hyper parameters! That would be helpful

[leffff]

hi，bro！have you solved this problem? Would you mind sharing the parameters for training 1D images？

So well I managed to start 1D U-Net

model = UNetModel(
    in_channels=198, # should be equal to num_features (input features) 
    dims=1, #this states, that we are using 1D U-Net
    model_channels=256, # inner model features
    out_channels=198, # should be equal to num_features (input features) 
    num_res_blocks=10, # idk
    attention_resolutions=("16",) # idk
)

x_0 = torch.rand(23, 64, 198) # our input [batch_size, num_atoms, num_features]
#num_atoms should be a 2 to some power
t = torch.rand(23) # our time [batch_size]

model(
    x=x_0, 
    timesteps=t, 
).shape # torch.Size([23, 64, 198]), which matches x.shape torch.Size([23, 64, 198])

[zjgt]

Hi, leffff, could you share more about your codes? I also started training with black and white images. I got a write_line error with the default hyperparameters. Also, how is the diffusion model work compared to StyleGAN?