Passing a small input sequence

[Sinjini15]

For the tutorial on the repo page if I try to pass

model = AudioDiffusionModel(in_channels=1)
x = torch.randn(2, 1, 2 ** 6)
loss = model(x)

I get the following error:

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 2 but got size 1 for tensor number 1 in the list.

Seems like for anything less than 2 ** 14 I get this error. Is there any way to pass a smaller input sequence to the diffusion model?

[flavioschneider]

You have to reduce the network patch_size and downsampling factors. Note however that 2**6 is a length of only 64, I doubt you can listen to anything meaningful with so few numbers (even very bad quality music has 6000 to 12000 samples per second). If you are doing something else that is not audio, you might want to do diffusion on something else that is not a UNet1d for such a short length.

patch_size=16, # This divides the input by 2**4
multipliers=[1, 2, 4, 4, 4, 4, 4], # (len=num_layers+1)
factors=[4, 4, 4, 2, 2, 2], # This divides the input the respective numbers (len=num_layers)
num_blocks=[2, 2, 2, 2, 2, 2], #  (len=num_layers)
attentions=[False, False, False, True, True, True], #  (len=num_layers)

[Sinjini15]

Thanks for explaining! I am just playing around with some other data, not specifically music :)

My data size is actually 2**9 but I was testing with 64 length data to see if it would even work. Your suggestion worked!

Posting for other people who might want to implement the same:

I reduced patch_size to 8 and factors to [2, 2, 2, 1, 1, 1] and it worked for me!