Upload musemorphic/tokenizer.py
Browse files- musemorphic/tokenizer.py +521 -0
musemorphic/tokenizer.py
ADDED
|
@@ -0,0 +1,521 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
MuseMorphic MIDI Tokenizer
|
| 3 |
+
==========================
|
| 4 |
+
|
| 5 |
+
REMI+ tokenization with BPE compression for MIDI files.
|
| 6 |
+
Handles: bar boundaries, positions, pitches, velocities, durations,
|
| 7 |
+
tempo, time signatures, instruments, and control attributes.
|
| 8 |
+
|
| 9 |
+
Based on:
|
| 10 |
+
- REMI (Huang & Yang, 2020) - Beat-aware positional encoding
|
| 11 |
+
- REMI+ (von Rütte et al., 2023) - Multi-track extension
|
| 12 |
+
- MIDI-RWKV (2025) - BPE compression for MIDI tokens
|
| 13 |
+
- MIDI-GPT (2025) - Attribute control tokens
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
import json
|
| 17 |
+
import math
|
| 18 |
+
import os
|
| 19 |
+
from dataclasses import dataclass, field
|
| 20 |
+
from typing import List, Dict, Tuple, Optional, Union
|
| 21 |
+
from pathlib import Path
|
| 22 |
+
|
| 23 |
+
import numpy as np
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
@dataclass
|
| 27 |
+
class TokenizerConfig:
|
| 28 |
+
"""Configuration for REMI+ tokenizer."""
|
| 29 |
+
|
| 30 |
+
# Resolution
|
| 31 |
+
ticks_per_beat: int = 480
|
| 32 |
+
max_bar_length: int = 4 # Max bars per phrase
|
| 33 |
+
position_resolution: int = 16 # 16th note grid
|
| 34 |
+
|
| 35 |
+
# Pitch
|
| 36 |
+
pitch_range: Tuple[int, int] = (21, 108) # Piano range (A0-C8)
|
| 37 |
+
|
| 38 |
+
# Velocity
|
| 39 |
+
n_velocity_bins: int = 32
|
| 40 |
+
|
| 41 |
+
# Duration
|
| 42 |
+
n_duration_bins: int = 64 # From 1/32 note to 8 whole notes
|
| 43 |
+
|
| 44 |
+
# Tempo
|
| 45 |
+
tempo_range: Tuple[int, int] = (30, 210)
|
| 46 |
+
tempo_step: int = 4
|
| 47 |
+
|
| 48 |
+
# Time Signature
|
| 49 |
+
time_signatures: List[Tuple[int, int]] = field(default_factory=lambda: [
|
| 50 |
+
(2, 4), (3, 4), (4, 4), (5, 4), (6, 4), (3, 8), (6, 8), (12, 8)
|
| 51 |
+
])
|
| 52 |
+
|
| 53 |
+
# Instruments (General MIDI programs to track)
|
| 54 |
+
max_tracks: int = 16
|
| 55 |
+
|
| 56 |
+
# BPE
|
| 57 |
+
bpe_vocab_size: int = 8192
|
| 58 |
+
|
| 59 |
+
# Special tokens
|
| 60 |
+
pad_token: str = "<PAD>"
|
| 61 |
+
bos_token: str = "<BOS>"
|
| 62 |
+
eos_token: str = "<EOS>"
|
| 63 |
+
mask_token: str = "<MASK>"
|
| 64 |
+
bar_token: str = "<BAR>"
|
| 65 |
+
track_start_token: str = "<TRACK_START>"
|
| 66 |
+
track_end_token: str = "<TRACK_END>"
|
| 67 |
+
phrase_start_token: str = "<PHRASE_START>"
|
| 68 |
+
phrase_end_token: str = "<PHRASE_END>"
|
| 69 |
+
|
| 70 |
+
# Control tokens
|
| 71 |
+
ctrl_density_prefix: str = "DENSITY"
|
| 72 |
+
ctrl_polyphony_prefix: str = "POLY"
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
class REMIPlusTokenizer:
|
| 76 |
+
"""
|
| 77 |
+
REMI+ tokenizer for MIDI files.
|
| 78 |
+
|
| 79 |
+
Converts MIDI → REMI+ token sequence → integer IDs.
|
| 80 |
+
Supports phrase-level segmentation for PhraseVAE.
|
| 81 |
+
|
| 82 |
+
Vocabulary structure:
|
| 83 |
+
[0] PAD
|
| 84 |
+
[1] BOS
|
| 85 |
+
[2] EOS
|
| 86 |
+
[3] MASK
|
| 87 |
+
[4] BAR
|
| 88 |
+
[5] TRACK_START
|
| 89 |
+
[6] TRACK_END
|
| 90 |
+
[7] PHRASE_START
|
| 91 |
+
[8] PHRASE_END
|
| 92 |
+
[9-24] Position_0 to Position_15
|
| 93 |
+
[25-31] TimeSig tokens
|
| 94 |
+
[32-76] Tempo tokens (30-210, step 4)
|
| 95 |
+
[77-204] Pitch tokens (21-108)
|
| 96 |
+
[205-236] Velocity tokens (1-32)
|
| 97 |
+
[237-300] Duration tokens (1-64)
|
| 98 |
+
[301-428] Program tokens (0-127)
|
| 99 |
+
[429-438] Density control tokens (1-10)
|
| 100 |
+
[439-448] Polyphony control tokens (1-10)
|
| 101 |
+
[449+] BPE merge tokens
|
| 102 |
+
"""
|
| 103 |
+
|
| 104 |
+
def __init__(self, config: Optional[TokenizerConfig] = None):
|
| 105 |
+
self.config = config or TokenizerConfig()
|
| 106 |
+
self._build_vocabulary()
|
| 107 |
+
|
| 108 |
+
def _build_vocabulary(self):
|
| 109 |
+
"""Build the base vocabulary before BPE."""
|
| 110 |
+
self.token_to_id = {}
|
| 111 |
+
self.id_to_token = {}
|
| 112 |
+
idx = 0
|
| 113 |
+
|
| 114 |
+
# Special tokens
|
| 115 |
+
for tok in [self.config.pad_token, self.config.bos_token,
|
| 116 |
+
self.config.eos_token, self.config.mask_token,
|
| 117 |
+
self.config.bar_token, self.config.track_start_token,
|
| 118 |
+
self.config.track_end_token, self.config.phrase_start_token,
|
| 119 |
+
self.config.phrase_end_token]:
|
| 120 |
+
self.token_to_id[tok] = idx
|
| 121 |
+
self.id_to_token[idx] = tok
|
| 122 |
+
idx += 1
|
| 123 |
+
|
| 124 |
+
# Position tokens
|
| 125 |
+
for p in range(self.config.position_resolution):
|
| 126 |
+
tok = f"Position_{p}"
|
| 127 |
+
self.token_to_id[tok] = idx
|
| 128 |
+
self.id_to_token[idx] = tok
|
| 129 |
+
idx += 1
|
| 130 |
+
|
| 131 |
+
# Time signature tokens
|
| 132 |
+
for num, den in self.config.time_signatures:
|
| 133 |
+
tok = f"TimeSig_{num}/{den}"
|
| 134 |
+
self.token_to_id[tok] = idx
|
| 135 |
+
self.id_to_token[idx] = tok
|
| 136 |
+
idx += 1
|
| 137 |
+
|
| 138 |
+
# Tempo tokens
|
| 139 |
+
for bpm in range(self.config.tempo_range[0], self.config.tempo_range[1] + 1, self.config.tempo_step):
|
| 140 |
+
tok = f"Tempo_{bpm}"
|
| 141 |
+
self.token_to_id[tok] = idx
|
| 142 |
+
self.id_to_token[idx] = tok
|
| 143 |
+
idx += 1
|
| 144 |
+
|
| 145 |
+
# Pitch tokens
|
| 146 |
+
for p in range(self.config.pitch_range[0], self.config.pitch_range[1] + 1):
|
| 147 |
+
tok = f"Pitch_{p}"
|
| 148 |
+
self.token_to_id[tok] = idx
|
| 149 |
+
self.id_to_token[idx] = tok
|
| 150 |
+
idx += 1
|
| 151 |
+
|
| 152 |
+
# Velocity tokens
|
| 153 |
+
for v in range(1, self.config.n_velocity_bins + 1):
|
| 154 |
+
tok = f"Velocity_{v}"
|
| 155 |
+
self.token_to_id[tok] = idx
|
| 156 |
+
self.id_to_token[idx] = tok
|
| 157 |
+
idx += 1
|
| 158 |
+
|
| 159 |
+
# Duration tokens
|
| 160 |
+
for d in range(1, self.config.n_duration_bins + 1):
|
| 161 |
+
tok = f"Duration_{d}"
|
| 162 |
+
self.token_to_id[tok] = idx
|
| 163 |
+
self.id_to_token[idx] = tok
|
| 164 |
+
idx += 1
|
| 165 |
+
|
| 166 |
+
# Program tokens (GM instruments)
|
| 167 |
+
for prog in range(128):
|
| 168 |
+
tok = f"Program_{prog}"
|
| 169 |
+
self.token_to_id[tok] = idx
|
| 170 |
+
self.id_to_token[idx] = tok
|
| 171 |
+
idx += 1
|
| 172 |
+
|
| 173 |
+
# Control tokens: density levels
|
| 174 |
+
for level in range(1, 11):
|
| 175 |
+
tok = f"{self.config.ctrl_density_prefix}_{level}"
|
| 176 |
+
self.token_to_id[tok] = idx
|
| 177 |
+
self.id_to_token[idx] = tok
|
| 178 |
+
idx += 1
|
| 179 |
+
|
| 180 |
+
# Control tokens: polyphony levels
|
| 181 |
+
for level in range(1, 11):
|
| 182 |
+
tok = f"{self.config.ctrl_polyphony_prefix}_{level}"
|
| 183 |
+
self.token_to_id[tok] = idx
|
| 184 |
+
self.id_to_token[idx] = tok
|
| 185 |
+
idx += 1
|
| 186 |
+
|
| 187 |
+
self.base_vocab_size = idx
|
| 188 |
+
self.vocab_size = idx # Will grow with BPE
|
| 189 |
+
|
| 190 |
+
# Special token IDs
|
| 191 |
+
self.pad_id = self.token_to_id[self.config.pad_token]
|
| 192 |
+
self.bos_id = self.token_to_id[self.config.bos_token]
|
| 193 |
+
self.eos_id = self.token_to_id[self.config.eos_token]
|
| 194 |
+
self.mask_id = self.token_to_id[self.config.mask_token]
|
| 195 |
+
self.bar_id = self.token_to_id[self.config.bar_token]
|
| 196 |
+
|
| 197 |
+
def midi_to_remi_tokens(self, notes: List[Dict], tempo: float = 120.0,
|
| 198 |
+
time_sig: Tuple[int, int] = (4, 4)) -> List[str]:
|
| 199 |
+
"""
|
| 200 |
+
Convert a list of note events to REMI+ token strings.
|
| 201 |
+
|
| 202 |
+
Args:
|
| 203 |
+
notes: List of dicts with keys: pitch, start, duration, velocity, program
|
| 204 |
+
tempo: BPM
|
| 205 |
+
time_sig: (numerator, denominator)
|
| 206 |
+
Returns:
|
| 207 |
+
List of REMI+ token strings
|
| 208 |
+
"""
|
| 209 |
+
if not notes:
|
| 210 |
+
return []
|
| 211 |
+
|
| 212 |
+
# Sort by start time
|
| 213 |
+
notes = sorted(notes, key=lambda n: (n.get('start', 0), n.get('pitch', 0)))
|
| 214 |
+
|
| 215 |
+
# Compute bar and position info
|
| 216 |
+
tpb = self.config.ticks_per_beat
|
| 217 |
+
beats_per_bar = time_sig[0] * (4.0 / time_sig[1])
|
| 218 |
+
ticks_per_bar = int(tpb * beats_per_bar)
|
| 219 |
+
ticks_per_position = ticks_per_bar // self.config.position_resolution
|
| 220 |
+
|
| 221 |
+
tokens = []
|
| 222 |
+
current_bar = -1
|
| 223 |
+
|
| 224 |
+
# Add track metadata
|
| 225 |
+
tokens.append(self.config.track_start_token)
|
| 226 |
+
|
| 227 |
+
# Time signature
|
| 228 |
+
ts_tok = f"TimeSig_{time_sig[0]}/{time_sig[1]}"
|
| 229 |
+
if ts_tok in self.token_to_id:
|
| 230 |
+
tokens.append(ts_tok)
|
| 231 |
+
|
| 232 |
+
# Tempo
|
| 233 |
+
tempo_bin = round(tempo / self.config.tempo_step) * self.config.tempo_step
|
| 234 |
+
tempo_bin = max(self.config.tempo_range[0], min(self.config.tempo_range[1], tempo_bin))
|
| 235 |
+
tokens.append(f"Tempo_{tempo_bin}")
|
| 236 |
+
|
| 237 |
+
for note in notes:
|
| 238 |
+
start = note.get('start', 0)
|
| 239 |
+
pitch = note.get('pitch', 60)
|
| 240 |
+
duration = note.get('duration', tpb)
|
| 241 |
+
velocity = note.get('velocity', 80)
|
| 242 |
+
|
| 243 |
+
# Compute bar number
|
| 244 |
+
bar = int(start // ticks_per_bar)
|
| 245 |
+
|
| 246 |
+
# Add bar tokens for any new bars
|
| 247 |
+
while current_bar < bar:
|
| 248 |
+
current_bar += 1
|
| 249 |
+
tokens.append(self.config.bar_token)
|
| 250 |
+
|
| 251 |
+
# Position within bar (quantized to grid)
|
| 252 |
+
pos_in_bar = start % ticks_per_bar
|
| 253 |
+
position = min(
|
| 254 |
+
int(pos_in_bar / ticks_per_position),
|
| 255 |
+
self.config.position_resolution - 1
|
| 256 |
+
)
|
| 257 |
+
tokens.append(f"Position_{position}")
|
| 258 |
+
|
| 259 |
+
# Pitch
|
| 260 |
+
pitch = max(self.config.pitch_range[0], min(self.config.pitch_range[1], pitch))
|
| 261 |
+
tokens.append(f"Pitch_{pitch}")
|
| 262 |
+
|
| 263 |
+
# Velocity (bin)
|
| 264 |
+
vel_bin = max(1, min(self.config.n_velocity_bins,
|
| 265 |
+
int(velocity / 128 * self.config.n_velocity_bins) + 1))
|
| 266 |
+
tokens.append(f"Velocity_{vel_bin}")
|
| 267 |
+
|
| 268 |
+
# Duration (bin)
|
| 269 |
+
dur_bin = max(1, min(self.config.n_duration_bins,
|
| 270 |
+
int(duration / ticks_per_position)))
|
| 271 |
+
tokens.append(f"Duration_{dur_bin}")
|
| 272 |
+
|
| 273 |
+
tokens.append(self.config.track_end_token)
|
| 274 |
+
return tokens
|
| 275 |
+
|
| 276 |
+
def encode(self, tokens: List[str]) -> List[int]:
|
| 277 |
+
"""Convert token strings to integer IDs."""
|
| 278 |
+
ids = [self.bos_id]
|
| 279 |
+
for tok in tokens:
|
| 280 |
+
if tok in self.token_to_id:
|
| 281 |
+
ids.append(self.token_to_id[tok])
|
| 282 |
+
ids.append(self.eos_id)
|
| 283 |
+
return ids
|
| 284 |
+
|
| 285 |
+
def decode(self, ids: List[int]) -> List[str]:
|
| 286 |
+
"""Convert integer IDs to token strings."""
|
| 287 |
+
tokens = []
|
| 288 |
+
for id_ in ids:
|
| 289 |
+
if id_ in self.id_to_token:
|
| 290 |
+
tok = self.id_to_token[id_]
|
| 291 |
+
if tok not in [self.config.pad_token, self.config.bos_token, self.config.eos_token]:
|
| 292 |
+
tokens.append(tok)
|
| 293 |
+
return tokens
|
| 294 |
+
|
| 295 |
+
def segment_into_phrases(self, tokens: List[str], bars_per_phrase: int = 1) -> List[List[str]]:
|
| 296 |
+
"""
|
| 297 |
+
Segment a full REMI+ token sequence into phrase-level chunks.
|
| 298 |
+
|
| 299 |
+
Each phrase = one bar of one track (following PhraseVAE convention).
|
| 300 |
+
"""
|
| 301 |
+
phrases = []
|
| 302 |
+
current_phrase = []
|
| 303 |
+
bar_count = 0
|
| 304 |
+
|
| 305 |
+
for tok in tokens:
|
| 306 |
+
if tok == self.config.bar_token:
|
| 307 |
+
bar_count += 1
|
| 308 |
+
if bar_count > bars_per_phrase and current_phrase:
|
| 309 |
+
phrases.append(current_phrase)
|
| 310 |
+
current_phrase = []
|
| 311 |
+
bar_count = 1
|
| 312 |
+
current_phrase.append(tok)
|
| 313 |
+
|
| 314 |
+
if current_phrase:
|
| 315 |
+
phrases.append(current_phrase)
|
| 316 |
+
|
| 317 |
+
return phrases
|
| 318 |
+
|
| 319 |
+
def compute_controls(self, phrase_tokens: List[str]) -> Dict[str, int]:
|
| 320 |
+
"""
|
| 321 |
+
Compute control attributes from a phrase's tokens.
|
| 322 |
+
|
| 323 |
+
Controls:
|
| 324 |
+
- density: number of notes (binned 1-10)
|
| 325 |
+
- polyphony: max simultaneous notes at any position (binned 1-10)
|
| 326 |
+
"""
|
| 327 |
+
note_count = sum(1 for t in phrase_tokens if t.startswith("Pitch_"))
|
| 328 |
+
|
| 329 |
+
# Density: bin into 1-10
|
| 330 |
+
density = min(10, max(1, int(note_count / 3) + 1))
|
| 331 |
+
|
| 332 |
+
# Polyphony: count notes at same position
|
| 333 |
+
positions = {}
|
| 334 |
+
current_pos = 0
|
| 335 |
+
for tok in phrase_tokens:
|
| 336 |
+
if tok.startswith("Position_"):
|
| 337 |
+
current_pos = int(tok.split("_")[1])
|
| 338 |
+
elif tok.startswith("Pitch_"):
|
| 339 |
+
positions[current_pos] = positions.get(current_pos, 0) + 1
|
| 340 |
+
|
| 341 |
+
max_poly = max(positions.values()) if positions else 1
|
| 342 |
+
polyphony = min(10, max(1, max_poly))
|
| 343 |
+
|
| 344 |
+
return {'density': density, 'polyphony': polyphony}
|
| 345 |
+
|
| 346 |
+
def pad_sequence(self, ids: List[int], max_len: int) -> List[int]:
|
| 347 |
+
"""Pad or truncate to max_len."""
|
| 348 |
+
if len(ids) >= max_len:
|
| 349 |
+
return ids[:max_len]
|
| 350 |
+
return ids + [self.pad_id] * (max_len - len(ids))
|
| 351 |
+
|
| 352 |
+
def save(self, path: str):
|
| 353 |
+
"""Save tokenizer to directory."""
|
| 354 |
+
os.makedirs(path, exist_ok=True)
|
| 355 |
+
data = {
|
| 356 |
+
'token_to_id': self.token_to_id,
|
| 357 |
+
'config': {
|
| 358 |
+
'ticks_per_beat': self.config.ticks_per_beat,
|
| 359 |
+
'position_resolution': self.config.position_resolution,
|
| 360 |
+
'pitch_range': list(self.config.pitch_range),
|
| 361 |
+
'n_velocity_bins': self.config.n_velocity_bins,
|
| 362 |
+
'n_duration_bins': self.config.n_duration_bins,
|
| 363 |
+
'tempo_range': list(self.config.tempo_range),
|
| 364 |
+
'tempo_step': self.config.tempo_step,
|
| 365 |
+
'bpe_vocab_size': self.config.bpe_vocab_size,
|
| 366 |
+
}
|
| 367 |
+
}
|
| 368 |
+
with open(os.path.join(path, 'tokenizer.json'), 'w') as f:
|
| 369 |
+
json.dump(data, f, indent=2)
|
| 370 |
+
|
| 371 |
+
@classmethod
|
| 372 |
+
def load(cls, path: str) -> 'REMIPlusTokenizer':
|
| 373 |
+
"""Load tokenizer from directory."""
|
| 374 |
+
with open(os.path.join(path, 'tokenizer.json'), 'r') as f:
|
| 375 |
+
data = json.load(f)
|
| 376 |
+
|
| 377 |
+
config = TokenizerConfig(**data['config'])
|
| 378 |
+
tokenizer = cls(config)
|
| 379 |
+
tokenizer.token_to_id = data['token_to_id']
|
| 380 |
+
tokenizer.id_to_token = {int(k): v for k, v in
|
| 381 |
+
{v: k for k, v in data['token_to_id'].items()}.items()}
|
| 382 |
+
return tokenizer
|
| 383 |
+
|
| 384 |
+
def tokens_to_midi_notes(self, tokens: List[str], ticks_per_beat: int = 480) -> List[Dict]:
|
| 385 |
+
"""
|
| 386 |
+
Convert REMI+ tokens back to note events.
|
| 387 |
+
|
| 388 |
+
Returns list of dicts: {pitch, start, duration, velocity}
|
| 389 |
+
"""
|
| 390 |
+
notes = []
|
| 391 |
+
current_bar = -1
|
| 392 |
+
current_position = 0
|
| 393 |
+
current_tempo = 120
|
| 394 |
+
time_sig = (4, 4)
|
| 395 |
+
|
| 396 |
+
beats_per_bar = 4.0 # default
|
| 397 |
+
ticks_per_bar = ticks_per_beat * 4
|
| 398 |
+
ticks_per_position = ticks_per_bar // self.config.position_resolution
|
| 399 |
+
|
| 400 |
+
# Pending note attributes
|
| 401 |
+
pending_pitch = None
|
| 402 |
+
pending_velocity = None
|
| 403 |
+
|
| 404 |
+
for tok in tokens:
|
| 405 |
+
if tok.startswith("TimeSig_"):
|
| 406 |
+
parts = tok.split("_")[1].split("/")
|
| 407 |
+
time_sig = (int(parts[0]), int(parts[1]))
|
| 408 |
+
beats_per_bar = time_sig[0] * (4.0 / time_sig[1])
|
| 409 |
+
ticks_per_bar = int(ticks_per_beat * beats_per_bar)
|
| 410 |
+
ticks_per_position = ticks_per_bar // self.config.position_resolution
|
| 411 |
+
|
| 412 |
+
elif tok.startswith("Tempo_"):
|
| 413 |
+
current_tempo = int(tok.split("_")[1])
|
| 414 |
+
|
| 415 |
+
elif tok == self.config.bar_token:
|
| 416 |
+
current_bar += 1
|
| 417 |
+
|
| 418 |
+
elif tok.startswith("Position_"):
|
| 419 |
+
current_position = int(tok.split("_")[1])
|
| 420 |
+
|
| 421 |
+
elif tok.startswith("Pitch_"):
|
| 422 |
+
pending_pitch = int(tok.split("_")[1])
|
| 423 |
+
|
| 424 |
+
elif tok.startswith("Velocity_"):
|
| 425 |
+
pending_velocity = int(tok.split("_")[1])
|
| 426 |
+
|
| 427 |
+
elif tok.startswith("Duration_"):
|
| 428 |
+
if pending_pitch is not None:
|
| 429 |
+
dur_bin = int(tok.split("_")[1])
|
| 430 |
+
start = current_bar * ticks_per_bar + current_position * ticks_per_position
|
| 431 |
+
duration = dur_bin * ticks_per_position
|
| 432 |
+
velocity = int((pending_velocity or 16) / self.config.n_velocity_bins * 127)
|
| 433 |
+
|
| 434 |
+
notes.append({
|
| 435 |
+
'pitch': pending_pitch,
|
| 436 |
+
'start': max(0, start),
|
| 437 |
+
'duration': duration,
|
| 438 |
+
'velocity': min(127, max(1, velocity)),
|
| 439 |
+
})
|
| 440 |
+
pending_pitch = None
|
| 441 |
+
pending_velocity = None
|
| 442 |
+
|
| 443 |
+
return notes
|
| 444 |
+
|
| 445 |
+
|
| 446 |
+
# ============================================================================
|
| 447 |
+
# MIDI File I/O Utilities (using midiutil or pretty_midi)
|
| 448 |
+
# ============================================================================
|
| 449 |
+
|
| 450 |
+
def notes_to_midi_file(notes: List[Dict], output_path: str,
|
| 451 |
+
tempo: float = 120.0, ticks_per_beat: int = 480):
|
| 452 |
+
"""
|
| 453 |
+
Write note events to a MIDI file.
|
| 454 |
+
|
| 455 |
+
Uses midiutil for lightweight MIDI writing (no heavy dependencies).
|
| 456 |
+
"""
|
| 457 |
+
try:
|
| 458 |
+
from midiutil import MIDIFile
|
| 459 |
+
|
| 460 |
+
midi = MIDIFile(1, ticks_per_quarternote=ticks_per_beat)
|
| 461 |
+
midi.addTempo(0, 0, tempo)
|
| 462 |
+
|
| 463 |
+
for note in notes:
|
| 464 |
+
pitch = note['pitch']
|
| 465 |
+
start_beat = note['start'] / ticks_per_beat
|
| 466 |
+
duration_beat = note['duration'] / ticks_per_beat
|
| 467 |
+
velocity = note['velocity']
|
| 468 |
+
|
| 469 |
+
midi.addNote(0, 0, pitch, start_beat, duration_beat, velocity)
|
| 470 |
+
|
| 471 |
+
with open(output_path, 'wb') as f:
|
| 472 |
+
midi.writeFile(f)
|
| 473 |
+
|
| 474 |
+
return True
|
| 475 |
+
except ImportError:
|
| 476 |
+
print("midiutil not installed. Install with: pip install midiutil")
|
| 477 |
+
return False
|
| 478 |
+
|
| 479 |
+
|
| 480 |
+
def midi_file_to_notes(midi_path: str) -> Tuple[List[Dict], float, Tuple[int, int]]:
|
| 481 |
+
"""
|
| 482 |
+
Read a MIDI file and extract note events.
|
| 483 |
+
|
| 484 |
+
Returns: (notes, tempo, time_signature)
|
| 485 |
+
"""
|
| 486 |
+
try:
|
| 487 |
+
import pretty_midi
|
| 488 |
+
|
| 489 |
+
pm = pretty_midi.PrettyMIDI(midi_path)
|
| 490 |
+
tempo = pm.estimate_tempo()
|
| 491 |
+
|
| 492 |
+
# Get time signature
|
| 493 |
+
if pm.time_signature_changes:
|
| 494 |
+
ts = pm.time_signature_changes[0]
|
| 495 |
+
time_sig = (ts.numerator, ts.denominator)
|
| 496 |
+
else:
|
| 497 |
+
time_sig = (4, 4)
|
| 498 |
+
|
| 499 |
+
notes = []
|
| 500 |
+
tpb = 480 # Standard ticks per beat
|
| 501 |
+
|
| 502 |
+
for instrument in pm.instruments:
|
| 503 |
+
if instrument.is_drum:
|
| 504 |
+
continue
|
| 505 |
+
for note in instrument.notes:
|
| 506 |
+
start_ticks = int(note.start * tempo / 60.0 * tpb)
|
| 507 |
+
duration_ticks = int((note.end - note.start) * tempo / 60.0 * tpb)
|
| 508 |
+
|
| 509 |
+
notes.append({
|
| 510 |
+
'pitch': note.pitch,
|
| 511 |
+
'start': start_ticks,
|
| 512 |
+
'duration': max(1, duration_ticks),
|
| 513 |
+
'velocity': note.velocity,
|
| 514 |
+
'program': instrument.program,
|
| 515 |
+
})
|
| 516 |
+
|
| 517 |
+
return notes, tempo, time_sig
|
| 518 |
+
|
| 519 |
+
except ImportError:
|
| 520 |
+
print("pretty_midi not installed. Install with: pip install pretty_midi")
|
| 521 |
+
return [], 120.0, (4, 4)
|