Ok, more clearly: that's just the name for our next release of stuff, but it's not about April's Fools
2019-04-02 09214, 2019
reosarevok
It was just we were hoping to be done with it by the start of April maybe? Dunno, yvanzo named it :D
2019-04-02 09239, 2019
pristine__
ruaok: hi. Classes over!
2019-04-02 09221, 2019
amCap1712 joined the channel
2019-04-02 09231, 2019
slc has quit
2019-04-02 09241, 2019
reosarevok
zas: I see you're mostly maintaining the murdos userscript repo now? I sent a PR with a tiny silly script that CatQuest requested - guess it might be of use to someone else too
2019-04-02 09200, 2019
reosarevok
Fixed it according to eslint rules a bit but let me know if it needs more changes :)
2019-04-02 09216, 2019
dolina has quit
2019-04-02 09222, 2019
CatQuest is a bit tiny silly
2019-04-02 09232, 2019
dolina joined the channel
2019-04-02 09208, 2019
moinmoi joined the channel
2019-04-02 09259, 2019
ruaok
pristine__: hi! I'm out for drinks with one of our customers right now. I'll ping you when I am done, ok?
I wonder if we should generate the recommendations for users and store them in hdfs.
2019-04-02 09250, 2019
ruaok
And not just a playlist worth, but a pile.
2019-04-02 09223, 2019
pristine__
yeah. Storing a trained model has no use as such when we have to run CF all over agian
2019-04-02 09225, 2019
pristine__
again*
2019-04-02 09206, 2019
ruaok
And then we can use track-track relationships and a recommendation history to make daily playlists.
2019-04-02 09245, 2019
pristine__
what do you mean by track-track relationship?
2019-04-02 09217, 2019
ruaok
The results of the annoy gsoc project
2019-04-02 09256, 2019
ruaok
How acoustically similar two tracks are.
2019-04-02 09215, 2019
pristine__
okay. That integration would be possible after May, I guess.
2019-04-02 09244, 2019
pristine__
What should be the next step from here? When we have a minimal recommendation system working
2019-04-02 09226, 2019
ruaok
Let's get the CF stuff working well.
2019-04-02 09201, 2019
ruaok
We need to train it to produce good results. So we need tools to evaluate how well it is doing.
2019-04-02 09238, 2019
pristine__
Let's say users listen to a few songs after visiting their daily mix. would a few song generate new recommendations?
2019-04-02 09259, 2019
pristine__
yeah, the same thing. good results
2019-04-02 09215, 2019
pristine__
there are inbuilt functions to train a model, I am not sure what can we do to train the model well
2019-04-02 09200, 2019
pristine__
ruaok: Umm.....what I can do is test what happens when we start generating recommendations using the data we have till now, and then pick a user, add their daily history and check what kind of result it gives?
2019-04-02 09219, 2019
pristine__
do you want to suggest anything?
2019-04-02 09244, 2019
ruaok returns home
2019-04-02 09200, 2019
ruaok
so, this changes the flow of data...
2019-04-02 09231, 2019
ruaok
we use the CF algorithm to make find a pile of "recommended tracks". let's say 100 per user.
2019-04-02 09245, 2019
ruaok
then for a daily mix we might use 5 or 10 of these tracks.
2019-04-02 09203, 2019
aliamo joined the channel
2019-04-02 09228, 2019
ruaok
combining them with the acoustically similar tracks, we can find acoustically related tracks and make a playlist.
2019-04-02 09241, 2019
ruaok
then we record that playlist, so we know what we've served up to our users.
2019-04-02 09212, 2019
ruaok
then when we generate a playlist for day 2, we know what we can use and what we can't because we served it before.
2019-04-02 09230, 2019
aliamo has quit
2019-04-02 09252, 2019
pristine__
ruaok:
2019-04-02 09235, 2019
pristine__
So when these 100 tracks are consumed, we train model again with the history collected over the days when these 100 were being used?
2019-04-02 09216, 2019
pristine__
this way we have appreciable number of songs to train the models again, is that so ?
2019-04-02 09228, 2019
ruaok
we may not need to retrain the model, but re-run the users' data based on their recent listens an in effort to get new recommendations.
2019-04-02 09237, 2019
ruaok
yes, I think so.
2019-04-02 09201, 2019
ruaok
now, I'm not an expert on recommendations either. I'm logically trying to reason this out as we go along.
2019-04-02 09245, 2019
pristine__
ruaok: why not retrain? Taste of users may change in time.
2019-04-02 09217, 2019
pristine__
me too :) Thanks for moving it along. I am kinda stuck at one point
2019-04-02 09228, 2019
ruaok
i'm not saying that we won't retrain. the questions is how often do we retrain?
2019-04-02 09246, 2019
ruaok
you see how slow it is, right? we don't want to do it more than we need to.
2019-04-02 09202, 2019
ruaok
problem is that we won't know that until we're actually listening to the output.
2019-04-02 09242, 2019
ruaok
and likely we'll get angry at how much it sucks. and then we'll work collectively to remove some of the suck.
so, remember the script I asked you to write? that makes the output human readable?
2019-04-02 09216, 2019
ruaok
once we have that we can start sanity checking things.
2019-04-02 09256, 2019
ruaok
once they are "sane" (read: not excessively pumped full of suck) we should make them and all the gory stats of how the model was trained available on a web site.
2019-04-02 09207, 2019
ruaok
then we should tweak the settings and train different version.
2019-04-02 09243, 2019
ruaok
repeat until we've crowd sourced something we things is a good start.
2019-04-02 09230, 2019
ruaok
in the meanwhile we can start working on the other, more mechanical, bits of the process. (taking the CF data, combining it will the similarity data and making playlists)
2019-04-02 09202, 2019
ruaok
amCap1712: not doing that is a long way off. sadly.
2019-04-02 09203, 2019
culinko joined the channel
2019-04-02 09203, 2019
culinko has quit
2019-04-02 09203, 2019
culinko joined the channel
2019-04-02 09214, 2019
culinko
hey guys
2019-04-02 09225, 2019
culinko
i need to confirm some behavior in the release editor
2019-04-02 09212, 2019
culinko
it's driving me crazy because i don't know if something is behaving like before or if something got changed
2019-04-02 09201, 2019
culinko
currently, if you're editing a release and you click on the "edit" button right next to the artist field, it opens a bubble on the right
2019-04-02 09202, 2019
culinko
then, if you highlight the artist (left click in the "artist in musicbrainz:" field, hold the mouse button and highlight the whole entry but you end outside of the bubble on the left, the bubble will close
2019-04-02 09229, 2019
culinko
was it always like this before? i can clarify if it's not clear what i mean
2019-04-02 09232, 2019
pristine__
ruaok: okay :)
2019-04-02 09252, 2019
culinko
i noticed this a couple of days ago because every time i want to select the artist name and remove it, i usually end up outside of the bubble which means it's a muscle memory
2019-04-02 09242, 2019
travis-ci joined the channel
2019-04-02 09243, 2019
travis-ci
metabrainz/picard#4581 (master - 00b13fe : Laurent Monin): The build passed.
culinko: I cannot reproduce it with Firefox, which browser/platform do you use?
2019-04-02 09247, 2019
culinko
i'm on chrome (win 7)
2019-04-02 09223, 2019
yvanzo
ok, lemme check Chrome/Linux
2019-04-02 09237, 2019
yvanzo
culinko: IIUYC, you press the mouse button into the field under the label “Artist in MusicBrainz:”, and hold the mouse button pressed while moving the cursor to the left (thus selecting field content) until the cursor goes out of the bubble box, right?
2019-04-02 09259, 2019
culinko
yes
2019-04-02 09200, 2019
yvanzo
(I can still not reproduce it with Chrome/Linux.)
2019-04-02 09254, 2019
moinmoi has quit
2019-04-02 09207, 2019
akhilesh joined the channel
2019-04-02 09241, 2019
akhilesh
Mr_Monkey: Around?
2019-04-02 09252, 2019
HorusHorrendus has quit
2019-04-02 09240, 2019
bitmap
I can reproduce it in chrome on macos but I'm not sure it's a new issue
2019-04-02 09209, 2019
bitmap
(it applies to any bubble and doesn't matter where you click inside it or which direction you drag the cursor)
2019-04-02 09243, 2019
bitmap
haven't touched the bubble code in ages though so I assume it was always like this
2019-04-02 09251, 2019
bitmap
culinko: ^
2019-04-02 09254, 2019
Gazooo joined the channel
2019-04-02 09212, 2019
nelgin
Why are foreign keys not created in the replication database but they are on the standalone?
2019-04-02 09237, 2019
bitmap
in some cases they can cause FK constraint violations due to the way replication packets are applied
2019-04-02 09243, 2019
bitmap
nelgin: why do you need foreign keys?
2019-04-02 09214, 2019
reosarevok
If you plan to change the data, you shouldn't replicate (it'll break it anyway)
2019-04-02 09221, 2019
reosarevok
If you don't plan to change the data, you don't need the fks
2019-04-02 09225, 2019
bitmap
right, I'm just wondering what you need them for in case I can help you find a different solution to whatever you're trying to do
Also, all the open -list edits are required for converting collection, which we should do for the rest of the schema change stuff, so ideally we can check those between this week and next week or something?