#metabrainz

That's a great one!!

That's fun!

yvanzo: hi! a few more questions on sir reindex timings, do you remember how long does indexing url and release group take on master and v3.0.1?

mayhem: this is like 3 years old, but you and MB get a nice shout out from Matthew Hawn at 21:30: https://mtflabs.net/podcast030/

Oh he’s on our board haha. Still nice though!

aerozol: the year in music 1987 https://usercontent.irccloud-cdn.com/file/8ehOr...

moooin!

aerozol: that receipt is great!

lucifer: I worked out some things last night that we keep forgetting....

Ooh the '87 one is cool too

Wish I'd come up with the receipt. Maybe I can come up with it later 😂

MTF has great design sense in general, but sticking matthew in bright light and taking his picture up close? WHY???

#moody

#toodumbtounderstandlightingbasics

this happened when I gave a talk for them too.

huh that recipt thing is kinda fun

:D

aerozol: if you reinvent the ticketbot, please use a background image that shows some paper grain.

the crinkles are a great idea, but on perfect white background, it falls a bit short.

imagine it with an askii sort of lb logo infront

Ooh that's cool. Unfortunately I don't know if we can copy that one without it being bloody obvious

oh! what about warez info file looking?

you know those with the big ascii logo an "we blah blah an then something similar to this recipt but bodging the titles toexklude some letters, make some letters int onumbers (l33t) and z's instead of s's..

or maybe we contact them and see if they are interested in making an LB version.

you can always ask them if it's ok

just an idea i had

I still reckon look at opening up the 'explore' section, the community out there is doing so much

Endless cool things

Warez one is cool too. Someone has to make a little song that plays!

oh. also an anecdote about date in hover on: I had some releases a couple days ago, was adding them, it said "released 6 years ago" etc.. "ugh thought I, I have to idk, math or something, and when going to copy the text / seeing if there ewas any button or something, then, and only then discovering that there was a hover with the actual date (badly done.)

then? I had to remember the date (repeating the numbers to myself (17, 08, 09)), as I entered them into mb.

in conclusion ,it would have been 30 times easier if I could have just copied a line of text to pasteinto mb, or even just seend a date and write it in mb while looking

aerozol: yea? you liked my warez idea? :D

Yeah cool as

:D

oh. lol what about the obvious? superman scroling text

I could mock these up all day (but I won't cause I've already spent a lot of time on that)

superman/starwars/

iirc they both had that

aaw

When we do video ones that would be cool

nah. just gif

Ok, I gotta go to bed, and this is very distracting/fun hah. Have a great day all!

or just a still with the 10 ones there like a page

heh :D i forgot it's late there

happy november to you

natta aerozol

it's already the 1s right? it'll be the 2 soon for you

mayhem: oh. which things?

a classic machine learning mistake -- it didn't trigger for me because what we're doing doesn't quite strike me as machine learning, but I guess it does. We never limited the input from any one user.

and then you get: https://listenbrainz.org/playlist/d79e0e22-71ed...

Daft Punk seed, Rammstein on track 3.

I bet there is one person who has a playlist with metal and daft punk and that runs on repeat and we can see it here.

so, I wish for two things:

1. Limit each user to X data points for any given entity-entity pair. I am guessing X should be low, like 3 or 5.

lucifer: there's a deprecation warning for frontend tests.. it's time for `actions/checkout@v3` https://usercontent.irccloud-cdn.com/file/twpLI...

2. (this might be tricky) is to get a top X users and their individual listen counts that contribute to each entity-entity pair.

so for 2, we would see: MBID0, MBID1, score, score_hist: {'user1": 1, "user2":76, "user3": 4 }

so that we can see which users contribute to the questionable rec.

this would seriously balloon the dataset, I totally get that.

but if we can implement this, we should delete all data sets and then re-run something like session_based_days_730_session_400_threshold_5_limit_200 which clearly exhibits the problem.

BUT, realistically speaking, we may not need to do #2 if #1 drastically improves the results.

mayhem: for 2, yeah even if you get the top contributor for each pair, it'll be ~10M rows easily. will add 1. we do this for sitewide stats already, so that 1 user's listens do not dominate the entire site's artist/release/recording stats.

exactly, agreed 100%.

so, lets start with the #1 and then removing matches from the same artist credit.

chinmay: sure sounds good to update those actions.

then start over and nuke all the datasets and I'll start generating new ones.

and I will actually code the troi patch to compare the matches against spotify audio features so I can run automated tests against data sets.

sounds good

boo that also seems like recomendations will all be too "narrow" too dough.. :/

dough?

though

∂ough

damn where is that thorn whench you need it

Compose key + t + h → þ

Compose key + T + H → Þ

Compose key is great 🥰

galactic0205: best to repeat your question in this channel as it is where the developers hang out

Hey everyone this is Prajwal Uday from Bangalore, India.

as a brief introduction of myself, I work on ML and NLP-related projects and have been researching algorithms and topics based on them.

So my query is, are there any projects related to Machine Learning to which I can contribute to? I'll be grateful if someone helps me out.

Thank you.

hi galactic0205

we don't do much NLP stuff, but we do do some machine learning using Spark and some homegrown stuff that may or may not qualify for "machine learning"

(we're using machine learning concepts, but only SQL, so can it be machine learning? lol)

I think Postgres will achieve sentience in a few years

oh that's alright

we're currently engaging in a project where were are trying to use the stream of listens for a user and trying to derive how individual recordings might relate to each other.

atj: in lucifer's hands, it will.

galactic0205: there is a project that I would like someone to do. its a nice and isolated project that doesn't depend on a lot of stuff.

namely, inferring genres.

that sounds great

MusicBrainz has a few million genres. but we have 24M or so tracks.

so the genre coverage is really not very good at all. not good enough to use for recommendations.