Update blocks.py

[ksasso1028]

Allow Unet to process any arbitrary size (above kernel size). Just need to pad a few samples :)

Model should work with any sample length, as opposed to just with powers of 2 @zqevans

[ksasso1028]

upon further inspection, @zqevans issue seems to be because you use a even kernel size for your fourier kernels in down/upscale, which means you arent padding enough using the typical padding formula for odd kernels ->(kernel - 1) // 2 * dil.

sample-generator/audio_diffusion/blocks.py

Line 71 in 6a3402d

[1 / 8, 3 / 8, 3 / 8, 1 / 8],

[ksasso1028]

updated: removed the concat by accident on upload for git, have fixed

[opasquetdotfr]

Hello!

How and where would you decrease the "kernel size"?
Is there a way to train and inference using smaller arrays blocks than what is currently is?

Thanks!

O.