---

name: deepjazz description: "Use deepjazz for legacy MIDI jazz generation with music21 preprocessing, abstract grammar tokens, Keras/Theano LSTM training, safe diagnostics, and modernization guidance." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

deepjazz Repo Skill

Use this skill when a task involves the deepjazz repository or a deepjazz-style workflow: symbolic MIDI jazz generation, Metheny-style MIDI preprocessing, abstract grammar tokens, note cleanup, Keras/Theano LSTM generation, optional Theano GPU commands, or modernizing the legacy code.

This skill is self-contained. Do not rely on reading original repo scripts at runtime to recover constants, signatures, or workflow boundaries; use the bundled references and scripts below.

First route by task

User task	Route
Inspect whether a MIDI file can work with deepjazz, change melody/accompaniment parts, or debug corpus extraction	sub-skills/midi-preprocessing/
Explain or validate grammar tokens, run parse/unparse/cleanup smoke checks, or debug malformed generated grammar	sub-skills/grammar-and-qa/
Run, adapt, or port the LSTM generation path; avoid playback/training hazards; inspect model settings	sub-skills/lstm-generation/
Diagnose install/import/backend problems that cut across all workflows	references/troubleshooting.md
Check whether this skill matches a checkout/source revision	references/repo-provenance.md

Minimal safe start

1. Treat deepjazz as a legacy script collection, not a modern installable package.

2. For faithful execution, use a Python 2.7-era environment with NumPy/SciPy, music21, Keras 1.x, and Theano. Read references/dependency-environment.md before installing.

3. Set KERAS_BACKEND=theano before importing legacy Keras.

4. Run the bundled no-training environment check:

   python scripts/check_deepjazz_environment.py --expect-theano
   

5. Validate MIDI structure before training:

   python sub-skills/midi-preprocessing/scripts/inspect_midi_structure.py --midi-file <input.mid>
   

6. Validate grammar/QA behavior without training:

   python sub-skills/grammar-and-qa/scripts/grammar_roundtrip_smoke.py --seed 7
   

7. Only run full generation after environment, data, and grammar checks pass. The original generation flow trains an LSTM, attempts realtime MIDI playback, and writes MIDI output.

Key distilled facts

• Public source modules: generator, grammar, preprocess, qa, and lstm.
• Verified public signatures: generate(data_fn, out_fn, N_epochs), build_model(corpus, val_indices, max_len, N_epochs=128), get_musical_data(data_fn), get_corpus_data(abstract_grammars), parse_melody(fullMeasureNotes, fullMeasureChords), unparse_grammar(m1_grammar, m1_chords), prune_grammar(curr_grammar), prune_notes(curr_notes), clean_up_notes(curr_notes).
• Default generation settings: max_len=20, max_tries=1000, diversity=0.5, bpm=130, default epochs 128.
• The README documents CPU execution and an optional NVIDIA/Theano GPU command; CPU is the safe baseline for this skill.
• The default MIDI workflow is tailored to a Metheny fixture: melody part 5, accompaniment parts 0,1,6,7, and solo window offsets 476..548.

Shared references and scripts

• references/dependency-environment.md: public legacy dependency guidance and safe command sequence.
• references/troubleshooting.md: cross-cutting install, Python compatibility, data, playback, speed, and optional CUDA troubleshooting.
• references/repo-provenance.md: source commit, branch, evidence paths, and refresh baseline.
• references/repo-routing-metadata.json: structured router metadata used only if this skill is later imported into the managed repo-skills library.
• scripts/check_deepjazz_environment.py: safe dependency/backend/module-signature diagnostic; accepts optional --repo-root for inspecting a local deepjazz-style source tree without training.

Non-goals

• Do not treat this as a generic modern music-generation framework.
• Do not claim GPU verification unless a task explicitly verifies the legacy Theano CUDA path.
• Do not run full generation as a quick smoke test; use bundled no-training helpers first.
• Do not expect arbitrary MIDI files to work without adapting part and offset assumptions.