spiyer99.github.io
spiyer99.github.io copied to clipboard
neelriyer
Output wav file not as expected
Hi!, First of all thanks very much for sharing, was very interesting to read. I tried to recreate the outputs myself but for some reason the wav file that was created was just a bad version (probably reconstructed) of the content audio, with no resemblance of the style audio. I am running the code in Pycharm and haven't changed any of the actual code besides the hyper parameters. I have tried playing with content/style weights, numsteps and learning rate. None have produced an output with similar style to the style audio. Am I doing something wrong? Do you have examples of outputs that have worked? TIA!
Hey @spiyer99 Thanks for your reply. I was reading this article: https://spiyer99.github.io/Change-Audio-Pytorch/
I tried running it myself with the following hyper parameters:
Style Weight:10000, Content Weight:2, Epochs:500, Learning Rate:0.0001
Where does the actual style learning happen? Am I missing something?
Also, is it saving the weights of the network? I mean if I wanted to run a "test" could it be done in 1 epoch?
TIA
So the style learning happens in CNNModel. It doesn't save the weights automatically, but you can definitely add that in. To do a "test" you'll probably need more than 1 epoch.
Tbh though the output on my machine wasn't very good. This was just an experiment. Please go ahead and add your own changes in!
Here's a few links that might help you: https://github.com/alishdipani/Neural-Style-Transfer-Audio https://pytorch.org/tutorials/advanced/neural_style_tutorial.html