Some unclear details.

[ysig]

Hi,

your code seems interesting and I tried to run it on a personal example.
First you don't specify the torch version, though it seems to work with the latest one.
Secondly, does this model work better with more data?
I try to run it on style of about 30 min and target of a bit more than 4 min but I receive an out of memory error in the gpu.
I tried to investigate the code, but I don't see any batch size, that I could edit so I could get your code to work in my example.

Can you propose anything?
Thanks in advance!

[mazzzystar]

For memory error question, I think you can modify N_FFT in model.py to a larger number, such as 1024/2048, so the STFT results will be of half length.

As for "the longer the better", current method computes the "Gram matrix" for style, therefore the matrix size for 4min and 30min is the same, so only in the assumption that the gram matrix for 30min convey more texture than 4min, will the results be better. So I can't make sure "the longer the better".