jmp_music: send me a private message with your email and I'll invite you.
2020-05-08 12940, 2020
alastairp
a great first step would be to train a model with gaia to ensure that you know how the process works, and then to reproduce that process in scikit learn
2020-05-08 12927, 2020
jmp_music
alastairp: Ok thanks! I downloaded also the datasets via the link you provided to me a few days ago.
2020-05-08 12956, 2020
alastairp
in gaia we do a grid search with about 700 C/gamma values and feature permutations. There's a configuration file which lists these parameters (https://github.com/MTG/gaia/blob/master/src/bindi…) It would be good to have something similar in scikit learn, but it doesn't have to use this configuration file
2020-05-08 12926, 2020
alastairp
I understand that sklearn has a number of helper tools for grid search, so it seems like it would be a good idea to use that as much as possible
2020-05-08 12921, 2020
jmp_music
I could test the sklearn's GridSearch embedded algorithm as well as the RandomizedSearchCV too
2020-05-08 12934, 2020
jmp_music
to see which one could provide better results
2020-05-08 12921, 2020
alastairp
yeah, those were the ones that I was thinking of
2020-05-08 12950, 2020
jmp_music
I agree with you to start by training a similar model to gaia and compare its results. As I saw from the datasets you provided to me, the problem is a multilabel classification and now a multiclass one
2020-05-08 12954, 2020
jmp_music
not*
2020-05-08 12939, 2020
jmp_music
Each row of the dataset inludes its MBID, and some genres labeled to the track
2020-05-08 12950, 2020
jmp_music
the genres are from 1 to 30
2020-05-08 12908, 2020
alastairp
for those datasets, yes - but this isn't our only classification task. The main reason that I sent you those datasets so that you could download them and have a local copy of a few thousand .json files
and I don't see any specific behaviour for rhythm_beats_position
2020-05-08 12918, 2020
alastairp
that's a good question though, I wonder if we should remove this from the data before building the model... it seems like it has the potential to introduce bad training data
2020-05-08 12903, 2020
jmp_music
Let me check about it and the other features too. Maybe some of them could be dropped before the training process and thus speed up the training time
2020-05-08 12916, 2020
alastairp
good idea, but let's focus on that after we've reproduced the existing models
2020-05-08 12937, 2020
jmp_music
yes of course.
2020-05-08 12941, 2020
jmp_music
for the training process and during the labeling of the data all these classes should be included?
Yeah, I suspect zas and outsidecontext are the most likely to be involved
2020-05-08 12935, 2020
lazka
ah, there is a history with user names for builds, so phillipp wolfer I guess
2020-05-08 12916, 2020
reosarevok
That'd be outsidecontext then
2020-05-08 12920, 2020
reosarevok
What did you need? :)
2020-05-08 12900, 2020
lazka
outsidecontext, you copied some of my mutagen packages into the PPA but because I moved the mutagen tools from the python2 to python3 package you also need to copy the python2-mutagen variants
2020-05-08 12931, 2020
lazka
(some user emailed me about it)
2020-05-08 12959, 2020
lazka
I should have added a version conflict I guess, but didn't think of the copying to other PPAs case
2020-05-08 12932, 2020
adhawkins has quit
2020-05-08 12946, 2020
adhawkins joined the channel
2020-05-08 12946, 2020
reosarevok
I'll send him an email about it too, in case he misses this :)
lazka: do I get this right: the issue is a file conflict, if one has the python2 package installed?
2020-05-08 12931, 2020
lazka
outsidecontext, if you install the py3 one it tries to install tools owned by the py2 package. To work around this I added a py2 variant which doesn't include the tools
2020-05-08 12916, 2020
lazka
so, yes
2020-05-08 12920, 2020
shivam-kapila
ruaok: Even though the query is heavy but we tend to do too much in /user/<user_name> route.
2020-05-08 12942, 2020
shivam-kapila
We even call this heavy query twice. And the second time its totally unbound
