Playlist Builder [EXPERIMENTAL] #209
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Implements a new service and page for extending playlists with similar songs. Key features: - Fetch and display existing playlists from the database - Find similar songs based on the entire playlist's centroid vector - Adjustable similarity threshold slider (0.0-1.0) - Configurable max songs limit - Accept/Decline buttons for each suggested song - Dynamic re-search when songs are accepted (included in centroid) or declined (excluded from results) - Save extended playlist with original + accepted songs to media server - Real-time statistics display showing original, accepted, declined, and suggested song counts Files added: - app_extend_playlist.py: Backend blueprint with API endpoints - templates/extend_playlist.html: Frontend UI with interactive controls Files modified: - app.py: Registered extend_playlist_bp blueprint - templates/sidebar_navi.html: Added navigation link
…s8mGAn5UghuFemm5ydF3 Claude/add openrouter support 01 c fs8m g an5 ughu femm5yd f3
CONFIRMED: GPU (CUDA) is currently only used for Analysis, not Clustering RESEARCH FINDINGS: - GPU acceleration can provide 10-30x speedup for clustering tasks - KMeans: 10-50x faster, DBSCAN: 5-100x faster, PCA: 10-40x faster - Example: 5000 clustering runs that take 2-4 hours on CPU can complete in 5-15 minutes on GPU IMPLEMENTATION: Implemented GPU-accelerated clustering using RAPIDS cuML as an optional feature: New Features: - GPU-accelerated KMeans, DBSCAN, and PCA using RAPIDS cuML - Automatic fallback to CPU if GPU unavailable or encounters errors - New environment variable USE_GPU_CLUSTERING (default: false) - Maintains all existing clustering options and settings - Compatible with CUDA 12.2+ and NVIDIA GPUs Files Modified: 1. config.py: - Added USE_GPU_CLUSTERING configuration flag 2. tasks/clustering_gpu.py (NEW): - Created GPU clustering module with RAPIDS cuML implementations - GPU-accelerated classes: GPUKMeans, GPUDBSCAN, GPUPCA - CPU-only wrappers: GPUGaussianMixture, GPUSpectralClustering - Factory functions for model creation with GPU/CPU selection - Automatic GPU availability detection and graceful fallback 3. tasks/clustering_helper.py: - Imported GPU clustering module with fallback handling - Updated _perform_single_clustering_iteration to use GPU PCA when enabled - Modified _apply_clustering_model to support GPU clustering - Maintains full backward compatibility with CPU-only mode 4. Dockerfile: - Added cupy-cuda12x and cuml-cu12 installation for NVIDIA builds - Only installs GPU packages when BASE_IMAGE is nvidia/cuda - CPU builds remain unchanged and lightweight 5. deployment/.env.example: - Added USE_GPU_CLUSTERING configuration with documentation - Default: false (CPU only, backward compatible) 6. README.md: - Added "GPU Acceleration for Clustering" section - Documented performance improvements and usage instructions - Listed supported algorithms and compatibility requirements - Noted that GaussianMixture and SpectralClustering use CPU (no GPU version) Notes: - GPU clustering is OPTIONAL and disabled by default - CPU clustering remains the default for backward compatibility - All existing clustering parameters and settings are preserved - GaussianMixture and SpectralClustering always use CPU (no cuML implementation) - GPU usage: Analysis (ONNX inference) + Clustering (RAPIDS cuML, optional)
…Ju9JeTX1N4popdd3Kt6FoM
The bash -lc subshell prevented access to the BASE_IMAGE ARG, causing GPU packages (cupy, cuml) to never be installed. Changed to use 'set -ux;' pattern (matching base stage) which properly accesses Docker ARG variables in the current shell. This ensures cuML and cupy are installed when building with nvidia/cuda base images, enabling GPU-accelerated clustering.
The new docker image needed packages that where quite large, resulting in a very slow docker build
Feat/gpu clustering
Dealing with empty responses, sql handling, and how it retires. The pre-requirements also would fail silently, now that's explicit and it asks for a retry
Feat/gpu clustering
…ct play URL for Jellyfin - Added `album`, `song_artist`, and `album_artist` columns to the `score` table in PostgreSQL. - Updated `tasks/mediaserver_jellyfin.py` to extract `SongArtist` and preserve `OriginalAlbumArtist` from Jellyfin items. - Updated `tasks/analysis.py` and `app_helper.py` to persist the new metadata fields during track analysis. - Updated `app_sonic_fingerprint.py` to include the new metadata fields and construct a direct stream URL in the API response.
- Dockerfile: Conditionally install `voyager` only in CUDA/GPU images. `voyager` requires AVX instructions which may crash on old CPUs. The non-CUDA image will now skip `voyager` to ensure compatibility with older hardware (Celeron/ARM).
- tasks/artist_gmm_manager.py:
- Wrap `import voyager` in try-except block to handle missing library gracefully on old CPUs.
- Add `VOYAGER_AVAILABLE` checks to disable artist similarity features if the library is missing, preventing crashes.
- Add support for GPU-accelerated GMM using `cuml.mixture.GaussianMixture` if available, falling back to `sklearn` otherwise.
- tasks/voyager_manager.py:
- Wrap `import voyager` in try-except block.
- Add `VOYAGER_AVAILABLE` checks to `build_and_store_voyager_index` and `load_voyager_index_for_querying` to prevent crashes on non-AVX systems.
- Added Smart Filters UI to Extend Playlist page, allowing users to select seed tracks based on metadata rules (Album, Artist, BPM, etc.).
- Integrated a sticky Web Player to play songs directly from search results.
- Updated results to display "Song Artist" instead of "Album Artist".
- Backend updates to `app_extend_playlist.py`:
- Added `_build_filter_query` to construct SQL from filters.
- Added `_get_stream_url` to generate playback URLs for Jellyfin/Emby/Navidrome.
- Updated `extend_playlist_api` to accept filters and return enriched metadata (song_artist, stream_url).
- Updated `save_extended_playlist` to handle filter-based sources.
- Frontend updates to `templates/extend_playlist.html`:
- Implemented Filter Builder UI.
- Implemented Sticky Audio Player.
- Added Play buttons to result rows.
Kept uv-based Dockerfile with requirements files from feat/gpu-clustering 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
Feat/gpu clustering
New feature that lets you weight tracks from x1 to x1024 to control how much they influence similar song recommendations. Features: - Weight buttons cycle: x1→x2→x4→...→x512→x1024→x1 - Source Playlist Drawer shows your tracks with weight controls - Works for both Smart Filter→Extend AND Existing Playlist flows - Remove button drops tracks; Exclude button pushes away similar songs - Auto-recalculates recommendations 500ms after weight changes Backend: - Weighted centroid: sum(w_i * v_i) / sum(w_i) - At x1024, one track = 99.1% of the centroid influence - Server-side weight validation Also fixed: weights and include/exclude state now survive recalculations TL;DR: That one perfect track can now bully the algorithm into submission 💪
Contributor
Author
|
Hi @NeptuneHub ,
Would there be things you're interested in for the main project? |
Owner
|
Hi @rendyhd, thanks again for your great work! It would be very interesting if you want integrate these features into the main project:
Regarding Plex, full support would be ideal (playlists, analysis, etc.). From what I’ve seen, Plex doesn't allow downloading tracks for analysis, but if you’ve found a workaround, that would be excellent. Before we move forward from moving the PR from draft to an actual one, I just ask that:
Thanks again for your contribution, really appreciated! |
Contributor
Author
|
Thanks for your commetns @NeptuneHub. How do you currently test? Do you have all media servers running? Currently I only have experience with Plex and Jellyfin (and Music Assistant running via Jellyfin) btw, did you have a dev discord, or prefer everything via github. Really love this project |
Owner
|
I’m kindly hosted in the Discord Server of the Jellify app (an amazing iOS/android client for jellyfin) where I have my AudioMuse-AI channel: You can definitely find me there! About test:
So basically yes, you add to install all and test. Especially when you change the media server API you need to retest all. I’m trying to implement more test, but are mainly unit test and don’t has big chance to find if you call the media server bad. So here the secret is:
I would like to get a VM in cloud strong enough, also to do some integration test on ALL the music server. But we need one big enough and start to have a cost. Being an opensource project without funds I preferred to avoid. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a draft to share the playlist builder I made. I marked it as experimental because I've changed a lot to get it to work the way I wanted it, which is not necessarily in line with the core purpose of this project.
NOTE: If you want to test this, I'd recommend you make sure it runs on a new postgres volume, it is not compatible the production one
It's built on top of #195
Summary
New Playlist Builder page that combines song discovery via Smart Filters with playlist extension capabilities, plus an integrated web player for previewing songs and dark mode support.
New Features
Playlist Builder Page
A new unified interface at
/playlist_builderwith two tabs:Smart Filters Tab
Existing Playlist Tab
Web Player
Dark Mode
Playlist Saving
_instantsuffix (for Playlist Builder saves)Next step: I'm looking to add star ratings, but unfortunately Jellyfin doesn't support this. Might have to experiment with Navidrome
Other Improvements
stream_urlandsong_artistto API responsestitleandauthorfieldsScreenshots