Skip to content

DATASET.md

Latest commit

 

History

History
291 lines (206 loc) · 9.08 KB

DATASET.md

File metadata and controls

291 lines (206 loc) · 9.08 KB

PyDrums Dataset Documentation

Overview

PyDrums uses a comprehensive dataset of professional drum patterns to train AI models for pattern generation. This document provides detailed information about the current dataset, sources, and capabilities.

Current Dataset Statistics

  • Total Patterns: 268 professional drum patterns
  • Training Examples: 1,331 AI training pairs
  • Musical Styles: 17 different genres
  • Time Signatures: Multiple (4/4, 12/8, 3/4, etc.)
  • Pattern Format: 6-token notation system (x, -, R, _, [, ])

Detailed Style Breakdown

1. Funk (150 examples)

  • Characteristics: Syncopated rhythms, heavy use of ghost notes, emphasis on groove
  • Common Patterns: Off-beat hi-hats, snare on beats 2 and 4, syncopated kick patterns
  • Example Patterns: "AfroCub1", "Funk1", "Funk2"

2. Rock (135 examples)

  • Characteristics: Steady backbeat, driving rhythm, straightforward patterns
  • Common Patterns: Hi-hat on eighth notes, snare on 2 and 4, kick on 1 and 3
  • Example Patterns: "Rock1", "Rock2", various rock subdivisions

3. General (122 examples)

  • Characteristics: Versatile patterns suitable for multiple genres
  • Common Patterns: Basic drum set combinations, foundational rhythms
  • Usage: Fallback patterns, learning examples, versatile grooves

4. Disco (110 examples)

  • Characteristics: Four-on-the-floor kick pattern, steady hi-hat, danceable groove
  • Common Patterns: Kick on every quarter note, snare on 2 and 4, open hi-hat accents
  • Era: 1970s dance music, modern house influences

5. Reggae (105 examples)

  • Characteristics: One drop rhythm, emphasis on beat 3, laid-back feel
  • Common Patterns: Kick and snare on beat 3, rim shots, syncopated hi-hat
  • Variations: One drop, rockers, steppers

6. Jazz (90 examples)

  • Characteristics: Swing feel, brush techniques, subtle ghost notes
  • Common Patterns: Swung eighth notes, ride cymbal patterns, brush rolls
  • Sub-genres: Swing, bebop, Latin jazz influences

7. Ballad (90 examples)

  • Characteristics: Slow tempo, emotional expression, space between notes
  • Common Patterns: Simple kick and snare, emphasis on dynamics, minimal hi-hat
  • Usage: Slow songs, emotional moments, sparse arrangements

8. Pop (90 examples)

  • Characteristics: Radio-friendly, accessible rhythms, clear structure
  • Common Patterns: Straightforward backbeat, consistent hi-hat, commercial appeal
  • Era: Modern pop music, chart-friendly patterns

9. R&B (90 examples)

  • Characteristics: Groove-oriented, soul influences, sophisticated rhythms
  • Common Patterns: Syncopated patterns, ghost notes, pocket playing
  • Sub-genres: Classic soul, modern R&B, neo-soul

10. Latin (85 examples)

  • Characteristics: Complex polyrhythms, percussion influences, cultural authenticity
  • Common Patterns: Clave patterns, montuno rhythms, cross-rhythms
  • Sub-genres: Salsa, mambo, cha-cha, general Latin

11. Afro-Cuban (75 examples)

  • Characteristics: Traditional African-Cuban rhythms, complex polyrhythms
  • Common Patterns: Son clave, rumba patterns, traditional percussion translations
  • Cultural Context: Authentic Cuban musical traditions

12. Shuffle (45 examples)

  • Characteristics: Swung eighth note feel, blues influences
  • Common Patterns: Triplet-based subdivisions, blues shuffle, Texas shuffle
  • Usage: Blues, country, rock variations

13. Bossa Nova (45 examples)

  • Characteristics: Brazilian jazz influences, subtle Latin groove
  • Common Patterns: Gentle Latin feel, jazz harmony support, sophisticated rhythm
  • Origin: Brazilian music, João Gilberto influence

14. Blues (45 examples)

  • Characteristics: Traditional blues rhythms, 12-bar support
  • Common Patterns: Shuffle feel, straight eighth notes, call and response
  • Sub-genres: Delta blues, Chicago blues, modern blues

15. Waltz (24 examples)

  • Characteristics: 3/4 time signature, classical influences
  • Common Patterns: Strong beat 1, lighter beats 2 and 3, flowing feel
  • Usage: Classical music, folk dances, ballroom dancing

16. March (20 examples)

  • Characteristics: Military precision, steady tempo, ceremonial feel
  • Common Patterns: Strong downbeats, snare drum rolls, parade rhythm
  • Usage: Military ceremonies, parades, patriotic music

17. Tango (10 examples)

  • Characteristics: Argentinian tango rhythm, dramatic pauses
  • Common Patterns: Distinctive tango clave, staccato accents, passionate expression
  • Origin: Argentinian dance music

Data Sources

Primary Source: DrumMachinePatterns260

  • URL: https://github.com/stephenhandley/DrumMachinePatterns/blob/master/Sources/DrumMachinePatterns260/Patterns.json
  • Format: JSON with structured pattern data
  • Content: 268 professional drum machine patterns
  • Structure: Each pattern includes title, signature, length, and track data
  • Quality: High-quality, professionally curated patterns

Data Structure Example

{
  "title": "AfroCub1",
  "signature": "4/4",
  "length": 16,
  "tracks": {
    "ClosedHiHat": ["Note", "Rest", "Note", "Note", ...],
    "SnareDrum": ["Rest", "Rest", "Rest", "Rest", ...],
    "BassDrum": ["Note", "Rest", "Rest", "Rest", ...]
  }
}

Legacy Sources

  • stephenhandley/DrumMachinePatterns: Individual JSON files (API access issues)
  • montoyamoraga/drum-machine-patterns: Markdown format patterns (reference)

Pattern Notation System

PyDrums uses a 6-token notation system for representing drum patterns:

  • x = Hit/strike the drum (main accent)
  • - = Rest/silence (no sound)
  • R = Roll (extended sound, multiple rapid hits)
  • _ = Ghost note (quiet hit, subtle accent)
  • [ = Flam start (grace note before main hit)
  • ] = Flam end (completion of flam)

Drum Abbreviations

  • ch = Closed Hi-Hat
  • oh = Open Hi-Hat
  • sd = Snare Drum
  • bd = Bass Drum (Kick)
  • hh = Hi-Hat Pedal
  • cc = Crash Cymbal
  • rc = Ride Cymbal
  • ht = High Tom
  • mt = Mid Tom
  • lt = Low Tom
  • rs = Rim Shot
  • cb = Cowbell
  • tb = Tambourine
  • hc = Hand Clap

Pattern Format

ch: x-x-x-x-x-x-x-x-; sd: ----x-------x---; bd: x-----x-x-------

Training Data Generation

Each source pattern generates multiple training examples:

  1. Style-based prompts: "Create a [style] drum pattern"
  2. Action-based prompts: "Generate a [style] beat"
  3. Format-based prompts: "Make a [style] drum loop"
  4. Time signature prompts: "Create a [time_sig] [style] pattern"

Example training pair:

{
  "input": "Create a funk drum pattern",
  "output": "ch: x-x-x-x-x-x-x-x-; sd: ----x--x----x---; bd: x-----x---x-----",
  "style": "funk",
  "time_signature": "4/4",
  "source_pattern": "Funk1"
}

Adding New Data Sources

PyDrums supports multiple data sources. To add a new source:

1. JSON Format Requirements

{
  "patterns": [
    {
      "name": "PatternName",
      "style": "funk",
      "timeSignature": "4/4",
      "drums": {
        "BassDrum": ["Note", "Rest", "Note", "Rest"],
        "SnareDrum": ["Rest", "Note", "Rest", "Note"]
      }
    }
  ]
}

2. CLI Command

pydrums setup-data --additional-url "https://example.com/patterns.json" --additional-name "my_source"

3. Supported Formats

  • Format A: drums key with drum names
  • Format B: tracks key with drum names (DrumMachinePatterns260)
  • Custom: Extend DataLoader class for other formats

Performance Metrics

Generation Quality

  • Valid Pattern Rate: ~85% of generated patterns follow correct format
  • Style Accuracy: High correlation between requested style and output
  • Musical Quality: Based on professional drum machine patterns

Dataset Coverage

  • Time Signatures: 4/4 (majority), 12/8, 3/4, 6/8
  • Tempo Range: Patterns suitable for 60-180 BPM
  • Complexity: From simple 4-beat patterns to complex 16-beat patterns

Technical Implementation

Few-Shot Learning

  • Uses 3-5 relevant examples for each generation
  • Selects examples based on keyword matching and style hints
  • Fallback to random examples if no matches found

Pattern Validation

  • Checks for proper drum notation format
  • Validates drum abbreviations against known mapping
  • Ensures pattern contains both hits and rests

MIDI Conversion

  • Uses General MIDI drum mapping (channel 9)
  • Supports all 6 notation tokens
  • Configurable tempo, loop count, and resolution

Future Enhancements

Planned Features

  1. Real-time training: Add patterns from user feedback
  2. Style transfer: Convert patterns between styles
  3. Complexity control: Generate simple or complex variations
  4. Humanization: Add timing and velocity variations

Additional Data Sources

  1. Live recording data: Real drummer performances
  2. Genre-specific collections: Specialized pattern libraries
  3. User contributions: Community-driven pattern sharing
  4. Historical patterns: Classic drum machine patterns from vintage hardware

Last updated: July 13, 2025 Dataset version: v1.0 with DrumMachinePatterns260 integration