Am i able to update and keep current analysis.

[Bigmack3000]

Hey, I just saw the notes on the new release from yesterday. So am I able to update to the latest version and keep my current analysis? I know it still has the noavx version, which is great, but it seems like a new scan is required to update anyway? That's another 3 months I can't use this then.

[NeptuneHub]

The situation is the following:

• The Essentia Muscinn model has been replaced with the original Musicnn model. These two models are very similar, so a new analysis is not required.
• The LAION CLAP model (used for text search) has been replaced with a lighter model I created called DCLAP. This change does require a new analysis, because the model is different.

If you are already using AudioMuse-AI with CLAP disabled, then a new analysis is not necessary. Otherwise, you will need to run it again.

That said, if you previously avoided enabling CLAP because the analysis was too slow, the new DCLAP model is significantly faster, so it may now be worth enabling.

One additional detail: if you update but don’t need to analyze new songs, you should still be able to keep using your existing analysis (I haven’t fully tested this scenario, but it should work).

Recommended approach

You can handle the update and the new analysis in parallel:

1. Keep your current deployment running

   • You can either leave it as it is or update it to the new version.
   • Do not run a new analysis on this instance.
   • If you have a cron task or a Jellyfin scheduled task that automatically runs analysis, disable it.
   • This allows you to keep using the existing analyzed songs and the new features without interruption.

2. Deploy a second container

   • Run the new version in a separate container.
   • Perform the new full analysis there.

Using multiple workers

Also keep in mind that you can deploy multiple workers.

For example, if your homelab is slow but you have a faster desktop or laptop, you can run an additional worker there to speed up the analysis.

For reference, when testing I frequently wipe my database and run a full analysis to ensure everything works for new users. Right now I’m running analysis with:

• 1 worker on my homelab (Intel)
• 1 worker on my Mac Mini (ARM)
• 1 worker on my Raspberry Pi 5 (to test compatibility)
• 1 worker on my laptop with a GPU (to test the GPU image)

edit: by thew way before doing anything, I suggest to do a backup of the database

[dbqk]

I have a dumb question. What is the best way to backup the database? Where are the results of the analysis stored?

[NeptuneHub]

Everything is stored in the database, so a full dump of the database is enough as a backup.

[Bigmack3000]

is clap enabled or disabled by default? whatever is default is what i would have done on my first scan.

i do need to analyze new songs, so i assume I will have to enable this either way? how much faster is it? if the original scan took months, what should i expect from this new scan?

what do i need to adjust in the yaml to set up a second container?

thanks!

[NeptuneHub]

As default CLAP is enabled and give you the possibility to use the text search functionality.

The new DCLAP is sensible faster. I assume that if with the same collection you needed 3 months, now should be 1 months maybe less.

If you deploy a second worker on your desktop/laptop, you can really shrink the time.

I suppose you use docker compose and I’m not very experienced in it (I use Kubernetes where you just need to create all in a different namespace). I don’t want to give you wrong command because I never did.

I think that if you backup the old database, and you ask suggestions to AI based on your actual docker compose (like how to deploy a second one of this) you should find a way with some test.

I did all this work exactly for people that need to analyze huge collections with old HW, but yes, I understood that having to reanalyze from scratch is not the best. Unlucky there isn’t another way.

[Bigmack3000]

thanks, i'll look into it. I'll post an update on how I go about it. Before I do, do you foresee any other reasons a new analysis scan would be required soon? because I'll just wait if i would have to do it again in a few months.

[Bigmack3000]

oh, also, i just checked and my yaml currently has clap as false, but i wouldn't have changed that, so it must have been the default. so would my scan time actually increase if i set it to true on the new version?

[NeptuneHub]

Look with clap disabled you run only Musicnn analysis that is a model of 0.7 million parameter.

CLAP is around 80 million.

The the new DCLAP is around 7 million parameter.

If you took 3 months with only Musicnn, means that you have a mix of very old hw and very big music collection.

For me you can keep clap disabled and just update without any need of rescan because the new Musicnn model is very similar.

The better alternatives if you want to enable clap (so the new DCLAP model but the env var is still named the same) is by adding oneor more workers on additional computer like your desktop/laptotp/whatelse.

[NeptuneHub]

The machine learning model is the heart of this project and I continually do research to improve it. It’s not that I like to ask people to redo the entire analysis, but if a good improvement can be reached or an impacting bug is raised, there is this possibility.

Say that at the moment I don’t have new idea for that for the short term.