Different shapes of track.stems depending in DB(is_wav)

[carlthome]

After decoding, I ran into what feels like a bug (if not, sorry!):

mus = musdb.DB()
tracks = mus.load_mus_tracks()
print(tracks[0].stems.shape)

(5, 7552000, 2)

mus = musdb.DB(is_wav=True)
tracks = mus.load_mus_tracks('train')
print(tracks[0].stems.shape)

(7552000, 2)

Also, when is_wav=True, load_mus_tracks returns an empty list if subsets=''.

[faroit]

oops, indeed... thanks for finding this. In the end, I did not use the .stems property in this tool, so I haven't done proper unit tests for it. I will fix it today and make a new release.

[faroit]

Reproduced in d0b06b9

[faroit]

should be fixed in 2405d18

@carlthome can you please check if this works as expected?

[carlthome]

Shapes look right now. Thanks!

mus = musdb.DB(is_wav=True)
tracks = mus.load_mus_tracks()

still fails though (while it works if is_wav=False)

[faroit]

uh, can you have a look if the sample set produces the same problems? You can create this by running fetch.sh and decode.sh.

Also do the py.tests run?

[faroit]

just double check... your decoded musdb file structure looks like this, right?

[carlthome]

Yeah, that's the folder structure after using the Docker scripts for decoding.

I'll go over to using the mp4:s though as they're nicer. I personally would be fine with the is_wav option being removed (to make the codebase simpler). Feel free to close this issue!

[faroit]

we have some other participants using the docker decoded version because they either don't use python or cannot easily install ffmpeg (hello windows users without anaconda)...

Therefore it would still be great to know if there is a bug.

[faroit]

probably fixed in v.0.2.1