Dimension issue with conditioning

[anshumansinha16]

I am trying to run the 1D diffusion model with conditioning but getting the following error.

x = self.init_conv(x) in line https://github.com/lucidrains/denoising-diffusion-pytorch/blob/5ff2393c72a2a678535ac1c31779684552f18189/denoising_diffusion_pytorch/denoising_diffusion_pytorch_1d.py#L352

The input is torch.Size([64, 705]) i.e batch size is 64 and input is 704+1 (1 for the conditioning) but when I put this in the int_conv ; the error is

RuntimeError: Given groups=1, weight of size [64, 2, 7], expected input[1, 64, 705] to have 2 channels, but got 64 channels instead

As the init_conv is Conv1d(2, 64, kernel_size=(7,), stride=(1,), padding=(3,)) ; I wish to know why do we have input_channels = channels * (2 if self_condition else 1) ; Do we need to have the input condition as a separate channel in our case?

[Nikos-86]

I had experienced a similar error before. The issue is that for the 1D, each of the input tensor dimensions mean something. That is , [ "number of 1D training sequences", "number of channels in each sequence", "length of each sequence" ] . The error says that it was expecting a dimension of sample size 64, 2 channels and length 7. Instead it got, 1 sequence, 64 channels, 705 length ( [1, 64, 705] ) . Maybe a permutation to change the dimensions to [64, 1, 705] could help? Also, you might want to change your "channels" variable in the Unet to 1 ( that is where the Conv1d gets the number of channels from. In your case it is set to 2 ). Give it a try, it might work.