GUI version that lets you input a song and it will continue it for you about jukebox HOT 6 OPEN

@diffractometer, the OP wants to prime the model. The variation on the colab notebook from @SMarioMan in #40 adds that. And it's easy. @InterestingMuck, just go open that notebook and read. You'll figure it out. You can't run this thing on your home PC unless you have a specialized GPU (by specialized GPU, I don't mean a GTX 2080). You can barely run it on the P100 that Google Colab (might) enable for you. So, that's the way to go.

from jukebox.

However, the 1b version just does not give the results of the 5b version.

I don't doubt it. Still, I personally prefer working with local hardware rather than paying for cloud compute time or dealing with the instability of Colab's free offering, and so given that I have a graphics card which can handle 1b that's where I'm enjoying spending my time with jukebox. I think I'm not the only one who would rather work this way with 1b than with 5b on colab, but maybe I'm wrong about that.

For what it's worth, as a musician, I've found that 1b can certainly provide some interesting and inspiring ideas.

At first brush I wasn't able to get co-composition working in sample.py, so what I've done is copy the relevant code blocks from the Colab notebook into a few py files and I'm calling those from a python prompt. That's working pretty well for me, but it's far from a GUI of course!

from jukebox.

See #40 for a colab-based solution.

from jukebox.

This code was put out "as is" so my feeling is if you want a GUI some people are going to have to want to build it themselves

edit: I do see some people are approving pr's so I guess I'm wrong? Anywho check out colab https://colab.research.google.com/github/openai/jukebox/blob/master/jukebox/Interacting_with_Jukebox.ipynb it's as close as it gets now.

from jukebox.

You can't run this thing on your home PC unless you have a specialized GPU (by specialized GPU, I don't mean a GTX 2080). You can barely run it on the P100 that Google Colab (might) enable for you.

That depends on what you're trying to do with it. I have no problem using an RTX 2080 Super to give 1b_lyrics three 1-minute prompts and get three upsampled 3-minute audio files back, in the styles of 3 different genres/artists. That process takes about 12 hours. I can also run co-composition on 1b_lyrics up to a length of 64 seconds. As long as you aren't using 5b_lyrics, it seems there's lots of room for exploration on a consumer 2080.

from jukebox.

@anlexmatos - thanks for the benchmark info. This is very interesting to hear.

However, the 1b version just does not give the results of the 5b version. So, what’s the easy interface to the 5b version? Colab of course.

Colab notebooks are the “GUI” for this. (Again, you can run miniature versions of this on your local machine, but why bother when you can rent a powerful GPU from the Goog?)

Issue #40 has pointers to the best version so far. Go check it out!

from jukebox.