#metabrainz

/

      • alastairp
        one thing that we never looked at, and has been an idea for years is to actually break down these features
      • 2019-03-12 07131, 2019

      • alastairp
        there's a way in SVM to actually look at where the separators are
      • 2019-03-12 07141, 2019

      • alastairp
        and then we can look at the things that were mis-classified
      • 2019-03-12 07148, 2019

      • reosarevok
        I'd be curious to see how those *overlap* though. To see what genres, according to just musical features, are basically the same
      • 2019-03-12 07100, 2019

      • alastairp
        "this was misclassified as folk, and not country because values xyz were in these ranges"
      • 2019-03-12 07112, 2019

      • alastairp
        and you can say "oh yeah, I can see how these might have been confused"
      • 2019-03-12 07132, 2019

      • alastairp
      • 2019-03-12 07134, 2019

      • reosarevok
        (a bunch will probably just differ because of lyrics and whatnot, yet will look basically the same to a machine analysis)
      • 2019-03-12 07143, 2019

      • alastairp
        hophop/ska, country/blues, folk/easy listening
      • 2019-03-12 07110, 2019

      • alastairp
        and it comes down to schema too
      • 2019-03-12 07125, 2019

      • alastairp
        where some genre trees have child genres as the children of weird parent genres
      • 2019-03-12 07137, 2019

      • alastairp
        the beets tree (which we used for this experiment) suffers from that a bit
      • 2019-03-12 07153, 2019

      • reosarevok
        Do I understand correctly that easylistening is only a bit more similar to itself than ska to latin music? :D
      • 2019-03-12 07131, 2019

      • alastairp
        mmm, not really
      • 2019-03-12 07107, 2019

      • reosarevok
        Oh, ok, I guess it's most "easylistening is very easy to misidentify as something else"
      • 2019-03-12 07111, 2019

      • reosarevok
        *mostly
      • 2019-03-12 07112, 2019

      • alastairp
        yeah
      • 2019-03-12 07131, 2019

      • alastairp
        also known as "there's no such genre as easy listening"
      • 2019-03-12 07140, 2019

      • alastairp
        or "what even is genre"
      • 2019-03-12 07111, 2019

      • reosarevok
        And turns out machines are great at realising electronic music isn't country
      • 2019-03-12 07116, 2019

      • reosarevok
        Who'd thought :D
      • 2019-03-12 07147, 2019

      • pulkit6559
        alastairp: ohh, i see. so when do we start performing this check?, because upto some point while creating the dataset we wont have enough data
      • 2019-03-12 07131, 2019

      • pulkit6559
        after an evalution job is completed?
      • 2019-03-12 07130, 2019

      • reosarevok
        alastairp: have you also tried it with (eh, forgot the actual term, I swear I studied this!) multi-grouping?
      • 2019-03-12 07158, 2019

      • reosarevok
        So that a recording can actually have multiple genres, to try to detect which ones are closer etc?
      • 2019-03-12 07115, 2019

      • alastairp
        oh yeah, that's a huge limitation of what we currently do
      • 2019-03-12 07148, 2019

      • alastairp
        https://acousticbrainz.org/7026e97e-bd6f-4852-8e8… I mean, we do actually give probablities for all classes
      • 2019-03-12 07134, 2019

      • alastairp
        this is part of the wish to get sklearn into the system, which would give us a much better system to swap out different algorithms that may be better at doing this kind of thing
      • 2019-03-12 07102, 2019

      • alastairp
        but there are other ideas, like instead of doing 1 classifier for 15 classes, do 15 "yes"/"no" classifiers
      • 2019-03-12 07129, 2019

      • reosarevok
        Yeah, I remember studying that :D
      • 2019-03-12 07135, 2019

      • alastairp
        there was a participant in one of our challenges a few years ago who did subgenre classifiers
      • 2019-03-12 07141, 2019

      • reosarevok
        It does seem more relevant to how genre works
      • 2019-03-12 07154, 2019

      • alastairp
        so they had 1 to say "it's one of these 10 main genres", and then another for each genre to say "it's this subgenre"
      • 2019-03-12 07102, 2019

      • alastairp
        working out to combine them is a bit more complex though
      • 2019-03-12 07125, 2019

      • alastairp
        also, because a 600-class classifier is a bit shit
      • 2019-03-12 07141, 2019

      • reosarevok
        I mean
      • 2019-03-12 07151, 2019

      • reosarevok
        You could do it in two steps, but not limit it to one main genre
      • 2019-03-12 07151, 2019

      • alastairp
        also, because genre is a bit shit
      • 2019-03-12 07155, 2019

      • alastairp
        is that what you want to say?
      • 2019-03-12 07156, 2019

      • reosarevok
        haha. That it is
      • 2019-03-12 07103, 2019

      • amCap1712
        outsidecontext: once tagger is set up
      • 2019-03-12 07126, 2019

      • alastairp
        but yeah, lots of ideas, and if I had the time to get a better ML framework in place we could try many of them
      • 2019-03-12 07132, 2019

      • reosarevok
        But still. If you run it for each main genre, then pick any where yes P is bigger than X and then run the subgenre classifiers for *those* only, it's a bit less shit :)
      • 2019-03-12 07141, 2019

      • alastairp
        right
      • 2019-03-12 07104, 2019

      • alastairp
        the concern is, if the main classifier gets it really wrong, then the subclassifer is a bit of a moot point
      • 2019-03-12 07110, 2019

      • reosarevok
        Sure
      • 2019-03-12 07123, 2019

      • reosarevok
        But then you should realize by seeing the subclassifiers are all like "wait wat"
      • 2019-03-12 07131, 2019

      • alastairp
        actually getting an evaluation for a model is pretty cheap though... there's no reason to not run it on _all_ subclassifiers
      • 2019-03-12 07109, 2019

      • alastairp
        and then use that to say "these sub-cls really strongly thought it was a particular subgenre, let's see if the main genre one picks one of these ones as the top probability"
      • 2019-03-12 07125, 2019

      • reosarevok
        I feel the main problem for subclassifiers would just be how similar many would be, so that you'd be struggling to find enough data to differentiate them without overfitting
      • 2019-03-12 07117, 2019

      • alastairp
        and then working out what defines a subgenre, and if it makes sense to put it as a subgenre of genre x, given the features that are being used to discriminate it
      • 2019-03-12 07142, 2019

      • alastairp
        pulkit6559: that's not super important... maybe we might decide to do it after each class in a dataset has at least 10 example, or 100
      • 2019-03-12 07143, 2019

      • dhruvang joined the channel
      • 2019-03-12 07111, 2019

      • dhruvan has quit
      • 2019-03-12 07120, 2019

      • BrainzGit
        [musicbrainz-server] mwiencek opened pull request #971 (master…merge-linked-entities-json): Merge new linked_entities during the request https://github.com/metabrainz/musicbrainz-server/…
      • 2019-03-12 07121, 2019

      • dhruvang has quit
      • 2019-03-12 07108, 2019

      • ruaok
        pristine--: any PRs for me to review?
      • 2019-03-12 07110, 2019

      • gr0uch0mars
        amCap1712: I have talked to a work mate about out problem with Cover Arts. I have posted the answer with a diagram in the Android thread: https://community.metabrainz.org/t/android-app-ar…
      • 2019-03-12 07112, 2019

      • reosarevok
        bitmap: would you suggest me to rebase my latest PR on top of that and go back to using the releaseGroupType then?
      • 2019-03-12 07117, 2019

      • ruaok
        iliekcomputers: you about today?
      • 2019-03-12 07129, 2019

      • amCap1712
        ok i'll see it gr0uch0mars
      • 2019-03-12 07149, 2019

      • bitmap
        reosarevok: sure, that sounds good
      • 2019-03-12 07159, 2019

      • pristine--
      • 2019-03-12 07110, 2019

      • ruaok
        great, on it.
      • 2019-03-12 07121, 2019

      • ruaok
        do you have spotify, pristine--?
      • 2019-03-12 07124, 2019

      • pristine--
        Just reached room from lab. I will start with server and get back to you.
      • 2019-03-12 07143, 2019

      • pristine--
        Yes
      • 2019-03-12 07150, 2019

      • ruaok
        uh oh. too much time hacking on open source stuff?
      • 2019-03-12 07152, 2019

      • pristine--
        Not premium.
      • 2019-03-12 07115, 2019

      • pristine--
        Yes. I almost consume everyday in open source
      • 2019-03-12 07126, 2019

      • ruaok
        are you familiar with "daily mixes" ?
      • 2019-03-12 07129, 2019

      • pristine--
        Have bunked a hell lot of classes. Attendance is short.
      • 2019-03-12 07139, 2019

      • pristine--
        Ma'am told me today. Leave it.
      • 2019-03-12 07142, 2019

      • pristine--
        Yes.
      • 2019-03-12 07153, 2019

      • ruaok
        tsk tsk. uni is more important than open source. please stay on top of your classes!
      • 2019-03-12 07156, 2019

      • pristine--
        We should build playlist?
      • 2019-03-12 07120, 2019

      • pristine--
        Like daily mixes
      • 2019-03-12 07132, 2019

      • ruaok
        yes.
      • 2019-03-12 07155, 2019

      • ruaok
        because the features of "make me a playlist, LB" doesn't fit into our systems very well.
      • 2019-03-12 07114, 2019

      • ruaok
        however, making daily mixes for people is much easier.
      • 2019-03-12 07120, 2019

      • ruaok
        interstingly enough.
      • 2019-03-12 07131, 2019

      • ruaok
        we can re-use some code bits to make this happen.
      • 2019-03-12 07140, 2019

      • pristine--
        Daily mixes also require history of users, right?
      • 2019-03-12 07148, 2019

      • ruaok
        you know how the user reports get automatically refreshed when the user logged in the last month?
      • 2019-03-12 07151, 2019

      • pristine--
        I mean they are not just tope trending
      • 2019-03-12 07156, 2019

      • ruaok
        yes, correct.
      • 2019-03-12 07159, 2019

      • reosarevok
        speaking of
      • 2019-03-12 07100, 2019

      • reosarevok
        "Something went wrong, please try again later."
      • 2019-03-12 07104, 2019

      • reosarevok
        Spotify broken for others too?
      • 2019-03-12 07114, 2019

      • ruaok
        unlink and relink, reosarevok.
      • 2019-03-12 07121, 2019

      • reosarevok
        No, no, not in LB, in general
      • 2019-03-12 07123, 2019

      • pristine--
        No. Not about this automatically refreshed.
      • 2019-03-12 07137, 2019

      • ruaok
        reosarevok: relink the matrix then. man ld
      • 2019-03-12 07152, 2019

      • pristine--
        What is this last month thing.
      • 2019-03-12 07158, 2019

      • ruaok
        don't worry about it.
      • 2019-03-12 07111, 2019

      • ruaok
        I'll make it clear in a sec.
      • 2019-03-12 07132, 2019

      • ruaok
        so, if a user has enough listens and was active in a day, we can create a daily mix for the user.
      • 2019-03-12 07142, 2019

      • ruaok
        automatically at midnight or something like that.
      • 2019-03-12 07101, 2019

      • ruaok
        the main server requests it for the user and the mixes get created in a batch.
      • 2019-03-12 07117, 2019

      • ruaok
        and then they get sent back to the main server.
      • 2019-03-12 07126, 2019

      • ruaok
        and like the reports, the main server stores that info for the user.
      • 2019-03-12 07139, 2019

      • ruaok
        and if the user listens to their daily mix, we'll make another one for them tomorrow.
      • 2019-03-12 07146, 2019

      • ruaok
        if they don't we won't bother updating it.
      • 2019-03-12 07119, 2019

      • ruaok
        and I think we make daily mixes out of the collaborative filtering stuff you're building.
      • 2019-03-12 07123, 2019

      • ruaok
        with me so far?
      • 2019-03-12 07104, 2019

      • pristine--
        Yes
      • 2019-03-12 07130, 2019

      • ruaok
        so, I propose that this ought to be the plan for the summer.
      • 2019-03-12 07140, 2019

      • ruaok
        and thus the heart of your gsoc proposal.
      • 2019-03-12 07110, 2019

      • pristine--
        Okay.
      • 2019-03-12 07114, 2019

      • pristine--
        Sounds cool
      • 2019-03-12 07125, 2019

      • pristine--
        I use Google docs?
      • 2019-03-12 07148, 2019

      • ruaok
        and for a bonus, if you have time, we'll create discover weekly. 🤣
      • 2019-03-12 07104, 2019

      • ruaok
        yes, google docs is great for this.
      • 2019-03-12 07109, 2019

      • pristine--
        Lol. Sure.
      • 2019-03-12 07128, 2019

      • pristine--
        Have you ever got 9/20?
      • 2019-03-12 07140, 2019

      • ruaok
        Yep.
      • 2019-03-12 07154, 2019

      • pristine--
        Good to hear.😆
      • 2019-03-12 07158, 2019

      • ruaok
        LIke a 3 out of 20 on a calculus exam that I thought I mastered.
      • 2019-03-12 07118, 2019

      • pristine--
        Hahaha. The less you get the more you know.
      • 2019-03-12 07130, 2019

      • Mr_Monkey has had his share of bad grades
      • 2019-03-12 07140, 2019

      • pristine--
        We are almost ready with graphs I think?
      • 2019-03-12 07141, 2019

      • ayerhart has quit
      • 2019-03-12 07147, 2019

      • pristine--
        Mr_Monkey: heya!
      • 2019-03-12 07100, 2019

      • amCap1712
        gr0uch0mars: Thats fine but the problem is that there is no efficient way of using a viewmodel inside the activity. We need to pass in a context and that will result in memory leaks. So the repository should directly return the data to the acitvity.
      • 2019-03-12 07105, 2019

      • ruaok
        I went back and cleaned up all my bad grades, except calculus 1. I understood it, but the teacher didnt give any partial credit.
      • 2019-03-12 07114, 2019

      • ruaok
        I understood the material, so I let the D stand.
      • 2019-03-12 07133, 2019

      • ruaok
        pristine--: getting there with graphs, yes.
      • 2019-03-12 07145, 2019

      • pristine--
        Grades don't matter that much, I think.
      • 2019-03-12 07154, 2019

      • pristine--
        Yay! Graphs would look cool.
      • 2019-03-12 07155, 2019

      • ruaok
        congratulations!
      • 2019-03-12 07104, 2019

      • ruaok
        grades don't matter a whole lot.
      • 2019-03-12 07112, 2019

      • pristine--
        Congratulations?
      • 2019-03-12 07119, 2019

      • ruaok
        do you think iliekcomputers will be hired when he graduates?
      • 2019-03-12 07123, 2019

      • pombreda has quit
      • 2019-03-12 07133, 2019

      • ruaok
        ha, too late. he could screw up all his courses and that won't change much.
      • 2019-03-12 07139, 2019

      • ruaok
        aww, shit. I should not give him any ideas.
      • 2019-03-12 07147, 2019

      • ruaok
        he might end up like reosarevok's mum
      • 2019-03-12 07100, 2019

      • pristine--
        Lol
      • 2019-03-12 07109, 2019

      • ruaok
        cogratulations on realizing that grades don't matter.
      • 2019-03-12 07118, 2019

      • pristine--
        😅
      • 2019-03-12 07112, 2019

      • ayerhart joined the channel
      • 2019-03-12 07125, 2019

      • reosarevok
        I mean, he will, eventually :p
      • 2019-03-12 07106, 2019

      • ruaok can't imagine iliekcomputers being fat and sassy
      • 2019-03-12 07154, 2019

      • BrainzGit
        [listenbrainz-server] mayhem merged pull request #545 (master…spotify-refresh-token-frontend): Refresh the Spotify token for the front-end player https://github.com/metabrainz/listenbrainz-server…
      • 2019-03-12 07105, 2019

      • BrainzGit
        [listenbrainz-server] mayhem merged pull request #546 (master…nuke-empty-api-endpoint): Remove unused endpoint https://github.com/metabrainz/listenbrainz-server…
      • 2019-03-12 07123, 2019

      • iliekcomputers
        I can see how the UK has the brains to come up with brexit now
      • 2019-03-12 07125, 2019

      • iliekcomputers
        How does anyone ask people to come with prints of digital documents and then scan them 🥴
      • 2019-03-12 07128, 2019

      • iliekcomputers
        So intelligent
      • 2019-03-12 07118, 2019

      • iliekcomputers has pretty good grades btw
      • 2019-03-12 07123, 2019

      • iliekcomputers
        🤣