Analyzing local music? about deej-ai HOT 5 CLOSED

The code should still work for your local files. What error do you get? I did rewrite the code for generating playlists more efficiently but that is in the deej-ai.online-app project which is linked to Spotify as you say. There is a more efficient way of generating the vectors using the calc_mp3tovecs and calc_tfidf scripts in the train directory. They were written with Spotify preview mp3s in mind, but will work with any (flat) directory of mp3s (MP3ToVec.py will work with nested directories).

from deej-ai.

Not getting any error so far, other than the CLI not accepting "foreign" tracks when the web tool does. I was just looking to re-use part of the integration work that you did for Spotify :)

I'll give it a go, maybe it's easier than expected ^^

from deej-ai.

I'm not sure I understand. Can you give a more detailed example of what you are trying to do and what you mean by CLI / web tool in this context? Thanks!

from deej-ai.

Sure!

As described in the readme, I ran the following command: python MP3ToVec.py Pickles mp3tovec --scan c:/your_music_library (with an updated path of course). Now I can start the web tool by running python Deej-A.I.py Pickles mp3tovec.
At the top, there's an area where I can simply drag-and-drop an mp3 file, and it will generate a mix based on that song. That can be any mp3.

However, I'm trying to generate m3u playlists instead, because this allows me to use the created mix in Jellyfin. Therefore, I'm running the command python Deej-A.I.py Pickles mp3tovec --playlist playlist_outfile.m3u --inputsong startingsong.mp3.
The problem seems to be that "startingsong.mp3" has to be in the existing keys, because if I use a path that was not within the scanned directory from the first command, I get the following error:

nsongs 10  
36 MP3s
Outfile playlist: playlist_outfile.m3u
Input song selected: C:\Users\Chaphasilor\Desktop\test.mp3
Requested 10 songs
Traceback (most recent call last):
  File "C:\Users\Chaphasilor\Code\Deej-AI\Deej-A.I.py", line 491, in <module>
    tracks = make_playlist([input_song], size=n_songs + 1, noise=noise, lookback=lookback)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\Chaphasilor\Code\Deej-AI\Deej-A.I.py", line 208, in make_playlist
    similar = most_similar(positive=playlist[-lookback:], topn=max_tries, noise=noise)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\Chaphasilor\Code\Deej-AI\Deej-A.I.py", line 179, in most_similar
    mp3_vec_i = np.sum([mp3tovec[i] for i in positive] + [-mp3tovec[i] for i in negative], axis=0)
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\Chaphasilor\Code\Deej-AI\Deej-A.I.py", line 179, in <listcomp>
                        ~~~~~~~~^^^
KeyError: 'C:\\Users\\Chaphasilor\\Desktop\\test.mp3

If I use a path from C:\\your-music-library\, it all works fine.

As for what I'm trying to do, in the end:

1. Pass the Jellyfin ID of any song to Deej-AI
2. Have Deej-AI analyze the song and generate a playlist of similar tracks from my library
3. Output the Jellyfin IDs of all songs in the playlists, so that I can add them to the queue

That means I need to:

• generate the waveforms/vectors based on my entire library
• pass the Jellyfin ID to Deej-AI
• have Deej-AI output a list of Jellyfin IDs

Given that you've done something very similar for Spotify (plus the whole downloading previews part), I was wondering if there was a way to build on top of the Spotify stuff, or if I would need to implement my own solution that translates Jellyfin IDs to file paths, passes that to Deej-AI, and then translates the resulting file paths from the playlist file back to Jellyfin IDs. But that would be cumbersome, that's why I would like to use Jellyfin IDs all the way through, just like you did with Spotify IDs (I assume).

from deej-ai.

Sorry for the delay.

Yes, the JoinTheDots.py script is expecting you to give it an already analyzed mp3, and the only way of knowing that is to pass the "key". The DeejAI.py script that runs the Flask webapp has some extra code which analyzes any mp3 track on the fly and then builds the playlist based on this, so you could take this code. But, as I say, these scripts are a bit old and clunky. If you are going to do anything at all serious, I would take inspiration from the train/calc_mp3tovecs.py, train/calc_tfidf.py and deej-ai.online/backend/deejai.py scripts. You would need to use the first two to pre-calculate the vectors for your entire library and you could adapt the functions to analyze a track and create playlists etc from the latter. I regret not having made the deejai.py code a bit more modular by separating out the Spotify dependencies, but I haven't really had the time to do that.

Good luck!

from deej-ai.