How mtg/mtg-jamendo-dataset works
MTG-Jamendo is a research dataset and companion scripts for music auto-tagging, not a hosted application. Researchers start with metadata and download audio or precomputed features, then use prepared tag subsets and artist-separated splits to train or test the CNN baseline and inspect results. The repository also supports metadata filtering, tag statistics, and MediaEval baseline evaluation. Edges reflect the README, sampled calls, or SOURCE INDEX references; some documented capabilities have no inspected implementation excerpt, so their links are limited to README evidence.
Main entry points: commons.py Metadata reader · raw.tsv Dataset metadata · download.py Dataset downloader · data_loader.py Feature loader
Components 21
Metadata and subsets
- Dataset metadata
data/raw.tsv - Train/test metadata
data/splits
Preparation and statistics
- Metadata reader
scripts/commons.py - Tag merging
scripts/clean_tags.py - Category subsets
scripts/filter_category.py - Tag-list filtering
scripts/filter_subset.py - Top-tag filtering
scripts/filter_toptags.py - Artist-separated splits
scripts/data_split.py - Tag statistics
scripts/get_statistics.py
Audio and features
- Dataset downloader
scripts/download/download.py - MTG data host
- Audio and features
data/download - Mel-spectrograms
scripts/melspectrograms.py
Baseline experiments
- Baseline runner
scripts/baseline/main.py - Feature loader
scripts/baseline/data_loader.py - Train and test
scripts/baseline/solver.py - CNN classifier
scripts/baseline/model.py - Baseline results
results/ml4md2019
Evaluation
- MediaEval naive baseline
scripts/mediaeval/baseline_naive.py - MediaEval results
results/mediaeval
Other components
- Researcher
Connections 25
- Researcher → Dataset metadata: uses metadata
README.md - Dataset metadata → Metadata reader: reads TSV
README.md - Metadata reader → Tag merging: reads metadata
scripts/clean_tags.py - Metadata reader → Category subsets: loads tracks
scripts/filter_category.py - Metadata reader → Tag-list filtering: loads tracks
scripts/filter_subset.py - Metadata reader → Top-tag filtering: loads tracks
scripts/filter_toptags.py - Metadata reader → Tag statistics: loads tracks
scripts/get_statistics.py - Dataset metadata → Artist-separated splits: splits metadata
README.md - Artist-separated splits → Train/test metadata: writes splits
README.md - Category subsets → Dataset metadata: writes subset
scripts/filter_category.py - Tag-list filtering → Dataset metadata: writes subset
scripts/filter_subset.py - Top-tag filtering → Tag statistics: computes tag counts
scripts/filter_toptags.py - Researcher → Dataset downloader: requests files
README.md - Dataset downloader → MTG data host: downloads archives
README.md - Dataset downloader → Audio and features: stores archives
README.md - Researcher → Mel-spectrograms: computes features
README.md - Baseline runner → Feature loader: creates loaders
scripts/baseline/main.py - Feature loader → Train/test metadata: loads split labels
README.md - Feature loader → Audio and features: loads features
scripts/baseline/data_loader.py - Baseline runner → Train and test: starts training or test
scripts/baseline/main.py - Train and test → CNN classifier: calls classifier
scripts/baseline/solver.py - Train and test → Baseline results: reports metrics
- Researcher → MediaEval naive baseline: runs evaluation
README.md - MediaEval naive baseline → Tag statistics: computes tag rates
scripts/mediaeval/baseline_naive.py - MediaEval naive baseline → MediaEval results: writes predictions