Does this work on windows? about jukebox HOT 19 OPEN

I've managed to run the inference on windows

The main problem is the coupling with torch.distributed, which is not needed for the inference, so:

• I replaced all get_rank and get_world_size calls with stubs, since you are running it on a single node, your rank is always 0 and the world size is 1
• removed the barrier calls along the execution path, same reason
• commented out the logging/metrics in logger.py because of the reduce logic in there
• fixed the encoding problem in artist_genre_processor.py, you have to provide explicit encoding="utf-8" when opening the labels file
• downloaded the models manually (might be not necessary, GSUtil doesn't work on my machine for some reason)
• op=dist.ReduceOp.SUM > op=None

the code looks super clumsy now but works

from jukebox.

Not at the moment unfortunately, need to tidy up the code a bit. I think I'll come up with the windows version by the end of the week.

from jukebox.

Ive added a windows compatible version I have up and running w/ sampling. Updated the readme https://github.com/peterlazzarino/jukebox

from jukebox.

Correction, am now much closer than I was before to getting this running, if I succeed I'll post steps

from jukebox.

@peterlazzarino I had to manually download the models, but this is working for me. Thanks!

from jukebox.

On Windows it is possible to not have to download the packages manually if you have wget.exe in your PATH. The model download fails since it uses wget which is mostly used with linux. Windows binaries do exist.

from jukebox.

Ive been hammering away at getting this running on mac and windows and can't get it going. Seems it's linux only based on the code's coupling to distributed pytorch but I could be very wrong, huge noob in python / ML.

from jukebox.

I've arrived at the same place haha. Are you able to run samples and get audio out? My computer is suddenly having a RAM issue and now won't start so this whole experiment is on pause for me :(

from jukebox.

Yep, I've managed to get the audio with 1b_lyrics and n_samples=1
PC configuration is rtx2070 + i7 4790, 24Gb RAM

Took significantly more time than stated in the article, about 4 hours for 10 seconds of audio

from jukebox.

Understandable since the times in the article I believe are generated using a Titan X. Do you have a fork I could try?

from jukebox.

@btrude good catch, updated the readme

from jukebox.

How do you manually download the sample?
(sry not very technical here lol)

from jukebox.

hey ocuvox.

When you get it running it will print messages in the console, you should see Downloading from gce and using cuda True

it will then hit an error if you dont have the models, it will look something like this

['wget', '-q', '-O', 'C:\\Users\\peter/.cache\\jukebox-assets/models/5b/vqvae.pth.tar', 'https://storage.googleapis.com/jukebox-assets/models/5b/vqvae.pth.tar'] Traceback (most recent call last): File "jukebox/sample.py", line 220, in <module>

The web URL to storage.googleapis.com is the link you will paste into your browser to download the file it expects. The path that is on your local machine is where you will put it. The folder, like C:\\Users\\peter/.cache\\jukebox-assets/models/5b/ has already been created, you just need to go there and drop the file once it finishes.

These are large files, 2-10GB and there are 4 of them needed to run the 1b_lyrics model

from jukebox.

Does this support the latest python version?

from jukebox.

Thanks @peterlazzarino ! Had hard times to set it up but could make it run on windows in 1B_lyrics. I'm a total beginner, so I hope you people with good knowledge will go on and adapt the rest! Best

from jukebox.

@peterlazzarino Would it be possible to enable issues on the fork?

from jukebox.

@camjac251 enabled, thanks

from jukebox.

so the main repo should support inference for windows now
except for downloading the models, I'll fix that by the end of this week

from jukebox.

@gnhdnb great to hear!

from jukebox.