[Announce] Music Similarity DSTM mixer

[bakker_be]

If you're referring to the lack of an HTML page response, that is to be expected. The server is running on port 11000, but its job is to not respond with an HTML page but with a list of tracks to feed DTSM, which it does. Running in debug mode you'll see this. It's easier to review the messages if you pipe into a log file, if you're interested in the details.

Based on what I see in the image, you should be able to get Similarity mixes from DTSM (assuming you've configured the Music Similarity plugin settings).

Try as I might, I can't seem to redirect the DEBUG output to a file. AFAIK "./music-similarity.py -l DEBUG > musicsimilarity.log" should send all output to the musicsimilarity.log file. It does create the file, but all other output remains in the terminal window The only info in the file is:

 * Serving Flask app 'lib.app' (lazy loading)
 * Environment: production
   WARNING: This is a development server. Do not use it in a production deployment.
   Use a production WSGI server instead.
 * Debug mode: off

[cpd73]

Understood, completely. It's your tool for your use case that you've chosen to freely share. I was commenting on what I saw after I took a peek behind the curtains and started to understand how things were connected. I don't currently have the skills to fork what you've done thus far to extend Similarity into SmartMix 2.0. But maybe someone else does or maybe you do, if the idea of creating playlists or influencing playlists based off of specific Musly/Essentia metadata is interesting or exciting. And if not, no complaints here. This is freeware/donation-ware, and I accept everything that comes with such a plugin.

I'm more than happy to add such functionality, or have others submit pull requests, etc, to implement this. Just that its not the current focus. I love the DSTM feature, its how I mainly listen to my music now - lpay 1 track, or album, and let LMS continue to add more tracks based on that.

I'm not a math(s) major

Me neither! I really am just making this up as I go along!

I've also only had an intro to statistics course at university. If 1000 tracks are selected to build the model music style database and then the 4 highest or lowest are used to filter based on a comparison to the seed track, I wonder if we're running into a sampling bias (or some other statistical bias). The 1000 tracks get further split into genres/genre groups based on relative percentage to the existing catalog and that affects the model. I do understand Similarity is pulling from the entire music collection when selecting a song to play and not just the 1000 model tracks. Like you, I only use local files; I do not use any streaming services.

I know nothing about statistics, or indeed how Musly does its similarity or how Essentia matches against models, etc. I'm just trying to put bits together! My initial thinking was that Musly is pretty good at getting similar tracks, so then adding filtering by BPM, Key, etc. would help improve even more.

Right now Similarity is a bit of a black box to me and it's not clear how much of an effect any one lever/setting has on a mix, so I've chosen to disable most settings and start fairly wide open. Next, I may go to the opposite extreme and use a very narrow range of setting values. As I get a sense for what's happening, I'll know where to go to make adjustments.

I agree, hence the initial post asking for best default settings, etc.

I do fully understand that and appreciate the distinction. The closer to 0 a value, the less likely it is that type; the closer to 1 the more likely it is. The intro to statistics course taught me that much. I do admit that I don't recall everything from the class though.

Sorry, I didn't mean to sound condescending. You probably know more about the theory of this than I

[cpd73]

Try as I might, I can't seem to redirect the DEBUG output to a file. AFAIK "./music-similarity.py -l DEBUG > musicsimilarity.log" should send all output to the musicsimilarity.log file. It does create the file, but all other output remains in the terminal window

Try:

./music-similarity.py -l DEBUG 2>&1 | tee musicsimilarity.log

'2>&1' ensures standard-error messages are placed on the same output as standard-output. '|' pipes the output of the left hand side to the right hand side. 'tee' takes this and shows on terminal and places into 'musicsimilarity.log'

[jd68]

 cmd > filename

will redirect only the standard output to the file "filename" but not the errors.

To collect both you have to use

 cmd > filename 2>&1

But I suggest to use

cmd | tee -a filename

That way you will see all output (standard and error) on the terminal and it will be written into the file "filename".

Edit: Damn! I am too slow. :-) and wrong: "tee" reads only from standard input. :-(

[Roland0]

It was a fantastic tool that allowed us to create mixes based on a variety of attributes — just like the ones in the music-similarity database. It had slider bars that allowed us to set a range for various attributes and then build a playlist from the tracks that met that criteria.

Imagine I want to query the music-similarity database for all songs with these values and produce a playlist—that was SmartMix and that's what I think the purpose of those values would or could be.

• happy: > 0.9
• aggressive: 0.4 - 0.6
• danceable: >0.75
• electronic: >0.8
• dark: >0.5

You can basically do this with LMS Essentia (using either extgui4lms or the LMS Playlist Editor as a frontend)

if you are interested in some background info regarding musly/essentia/similarity, see this thread

[bakker_be]

Try:

./music-similarity.py -l DEBUG 2>&1 | tee musicsimilarity.log

'2>&1' ensures standard-error messages are placed on the same output as standard-output. '|' pipes the output of the left hand side to the right hand side. 'tee' takes this and shows on terminal and places into 'musicsimilarity.log'

That did the trick! Thanks, "tee" is now stored in tips & tricks toolbox

[cpd73]

Imagine I want to query the music-similarity database for all songs with these values and produce a playlist—that was SmartMix and that's what I think the purpose of those values would or could be.

• happy: > 0.9
• aggressive: 0.4 - 0.6
• danceable: >0.75
• electronic: >0.8
• dark: >0.5

That criteria in my music collection finds 124 songs. I'd like to put that on play.

Thinking about this some more. Would it make sense to allow creation of 'Smart Mixes' that will select a number of tracks, allow this to be added to the queue, and then DSTM takes over adding new tracks. DSTM, however, would not know about the SmartMix, just that these would be the initial songs to create mixes from. If so creating another API that returns X songs based upon some attributes should be doable. It then requires the plugin to implement a JSONRPC and the creation of a UI to create these mixes. Obviously for me the UI would be created in Material. I know SlimBrowse would allow some sort of UI creation for this, but that's beyond my knowledge level at the mo - and as I only use Material it's not of major importance to me.

e.g. I would add a 'Smart Mixes' (or similar) entry to 'My Music' to list current mixes, allow editing, etc. The edit dialog would allow you to specify which attributes to use, number of tracks, and (perhaps) order of tracks (closest to attributes, random, etc.)

[cpd73]

If 1000 tracks are selected to build the model music style database and then the 4 highest or lowest are used to filter based on a comparison to the seed track, I wonder if we're running into a sampling bias

The "4" that I referred to was not tracks, but attributes. As in I find the 4 strongest Essentia attributes a track has (so attributes >=0.8 or <=0.2) and then filter tracks against those. What my script does is:

for each seed track (the LMS plugin sends a maximum of 5)
    get 5000 most similar tracks for Musly
        for each of these similar tracks
            filter out based on meta-data - to stop repeated titles, artists, etc.
            check BPM of candidate track against seed track
            check key of candidate track against seed track
            check loundness of candidate track against seed track
            for each of the 4 strongest Essentia attributes of seed track
                check candidate track's attribute aginst seed track
        stop when we have enough matching tracks for this seed

randomise selected tracks
return required amount of tracks

...there's more to it than that, but that is the basic idea of the filtering.

[prabbit]

You can basically do this with LMS Essentia (using either extgui4lms or the LMS Playlist Editor as a frontend)

if you are interested in some background info regarding musly/essentia/similarity, see this thread

I tried to install your plugins and was successful with some of them, but others require either systems, knowledge, or skills that I don't have at the moment. I remember using extgui4lms many years ago when it was still very new. No doubt your contributions have been important and valuable, because they helped Craig with his Similarity project. If I could figure out how to take my existing high-level Essentia JSON analysis files and have them re-processed by a Windows-based or Ubuntu-on-Windows-built tool, then I'd probably be able to proceed.

Thinking about this some more. Would it make sense to allow creation of 'Smart Mixes' that will select a number of tracks, allow this to be added to the queue, and then DSTM takes over adding new tracks. DSTM, however, would not know about the SmartMix, just that these would be the initial songs to create mixes from. If so creating another API that returns X songs based upon some attributes should be doable. It then requires the plugin to implement a JSONRPC and the creation of a UI to create these mixes. Obviously for me the UI would be created in Material. I know SlimBrowse would allow some sort of UI creation for this, but that's beyond my knowledge level at the mo - and as I only use Material it's not of major importance to me.

e.g. I would add a 'Smart Mixes' (or similar) entry to 'My Music' to list current mixes, allow editing, etc. The edit dialog would allow you to specify which attributes to use, number of tracks, and (perhaps) order of tracks (closest to attributes, random, etc.)

This makes sense to me: Mood Mixes.

I also only use Material (in Chrome, Material app on Android, in SLX on Windows).

It seems straightforward enough to build the API and UI to expose all low- and high-level attributes in the music-similarity.db file; that is, if one knows how to do these sorts of things.
Ask users to adjust sliders to produce a range for each attribute, along with some onscreen hints that each attribute indicates the strength of the probability the song matches that attribute type.
Gather all responses, pass those to the database, return a result. (Maybe 'maximum number of files to return' is also an attribute to pass so you're not returning thousands of entries, or if you do it's because the user specifically requested it.)
The result could be a new playlist. I'd likely almost always choose 'random', but others may want them played in some order — of course, what order is unknown because no attribute would be weighted more than others, so there'd be no way to rank them. Unless, that too, was something that was requested: rank by [ordered list of attributes].
For the UI, I'd convert the attribute values to 0 to 100 rather than 0 to 1, since most humans understand that easier than 0 to 1. If a user doesn't want to filter on an attribute they should allow all values by having the slider range extend from 0 to 100 — the default position. Behind the scenes I would also only use the hundredths position (e.g., 0.91) and no further (e.g., 0.9152377), although rounding to 0.92 would be expected here. I wouldn't expect rounding to the hundredths position would materially (no pun intended) affect the mix.

After the playlist runs out, it seems reasonable that DTSM would kick in using whatever rules are set for DTSM.

[Roland0]

If I could figure out how to take my existing high-level Essentia JSON analysis files and have them re-processed by a Windows-based or Ubuntu-on-Windows-built tool, then I'd probably be able to proceed.

LMS Essentia analysis/upload should run on any Linux system, so Ubuntu-on-Windows should work (using the "Gaia/SVM" variant as described on the homepage)
Re-using the existing high-level Essentia JSONs should be possible, with one provision: LMS Essentia adds the file's path (as used by LMS) to the JSON (as metadata.tags.file_path), so the existing JSONs would have to be modified before uploading them to LMS Essentia's database.