Spaces:
Sleeping
Sleeping
File size: 16,147 Bytes
b57c46e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 | """
Multi-feature score graph dataset for LaM-SLidE autoencoder training.
Provides ScoreGraphMultiFeatureDataset, collate functions, and feature specs.
The dataset loads .pt graph files produced by process_dataset.py and extracts
configurable subsets of 14 per-note features for ablation studies.
Feature tensor: 14 columns in graph['note'].x (from process_dataset.py).
Col Feature Raw Range Transform Vocab Notes
0 grid_position [0, 32] none 33 tokenised (16th grid)
1 micro_offset [0, 12] none 13 tokenised micro-shift
2 measure_idx [0, 255] none 256 0-based bar index
3 voice [0, 24] none 25 already 0-based (20 unique)
4 pitch_step [0, 6] none 7 C D E F G A B
5 pitch_alter [0, 4] none 5 shifted +2 in extraction
6 pitch_octave [-3, 11] +3 15 shift to 0-base
7 duration [0, 726] none 727 tokenised duration vocab
8 clef [0, 5] none 6 treble=0 bass=1 ...
9 ts_beats [1, 80] remap 57 remap 57 observed values
10 ts_beat_type {1,2,4,8,16,32} remap 6 remap 6 sparse -> [0,5]
11 key_fifths [-7, +7] +7 15 circle of fifths
12 key_mode [0, 1] none 2 0=major, 1=minor (optional)
13 staff [0, 36] none 37 already 0-based
"""
from dataclasses import dataclass, field
from pathlib import Path
from typing import Dict, List, Optional
import numpy as np
import torch
from torch.utils.data import Dataset
@dataclass
class NoteFeatureSpec:
"""Specification for a single note feature."""
name: str
col_index: int # Column index in graph['note'].x tensor
vocab_size: int
shift: int = 0 # Value to add to shift negative values to positive
# Optional remapping for sparse value sets (e.g., ts_beat_type: 2,4,8,16,32 -> 0,1,2,3,4)
remap: Optional[Dict[int, int]] = None
# Remapping for ts_beat_type: actual denominators -> consecutive tokens
# After dropping files with beat types {3, 5, 6, 9, 64}: remaining {1, 2, 4, 8, 16, 32}
TS_BEAT_TYPE_REMAP = {1: 0, 2: 1, 4: 2, 8: 3, 16: 4, 32: 5}
# Remapping for ts_beats: observed numerator values -> consecutive tokens.
# 57 unique values found in filtered data (max_notes=1530, max_bars=256).
# Unknown values are clamped to the closest entry in __getitem__.
TS_BEATS_REMAP = {
1: 0, 2: 1, 3: 2, 4: 3, 5: 4, 6: 5, 7: 6, 8: 7, 9: 8, 10: 9,
11: 10, 12: 11, 13: 12, 14: 13, 15: 14, 16: 15, 17: 16, 18: 17, 19: 18, 20: 19,
21: 20, 22: 21, 23: 22, 24: 23, 25: 24, 26: 25, 27: 26, 28: 27, 29: 28, 30: 29,
31: 30, 33: 31, 34: 32, 35: 33, 36: 34, 37: 35, 39: 36, 41: 37, 42: 38, 45: 39,
47: 40, 48: 41, 54: 42, 55: 43, 56: 44, 57: 45, 58: 46, 59: 47, 60: 48, 61: 49,
62: 50, 63: 51, 64: 52, 67: 53, 69: 54, 73: 55, 80: 56,
}
# Column indices in graph['note'].x tensor (shape: [num_notes, 14])
# Matches FEATURE_COLUMNS in process_dataset.py (verified 2025-02-07)
COL_INDICES = {
'grid_position': 0, # tokenised grid position
'micro_offset': 1, # tokenised micro offset
'measure_idx': 2, # 0-based bar index
'voice': 3, # 0-based voice
'pitch_step': 4, # 0-6 (C...B)
'pitch_alter': 5, # 0-4 (shifted +2)
'pitch_octave': 6, # raw octave (-3..11)
'duration': 7, # tokenised duration
'clef': 8, # 0-5
'ts_beats': 9, # raw numerator
'ts_beat_type': 10, # raw denominator {2,4,8,16}
'key_fifths': 11, # circle of fifths (-7..+7)
'key_mode': 12, # 0=major, 1=minor
'staff': 13, # 0-based global staff
}
# Pre-defined feature specifications for the 14-column feature tensor.
# Updated 2026-02-10 for new pipeline (process_dataset.py, max_notes=1530, max_bars=256).
# Value ranges verified via data_utils/analyze_features.py on full 68,542 graphs.
FEATURE_SPECS = {
# Position features
'grid_position': NoteFeatureSpec('grid_position', col_index=0, vocab_size=33), # tokenised [0, 32]
'micro_offset': NoteFeatureSpec('micro_offset', col_index=1, vocab_size=21), # tokenised [0, 20]
# Bar index
'measure_idx': NoteFeatureSpec('measure_idx', col_index=2, vocab_size=256), # 0-based [0, 255]
# Voice and staff (already 0-based from extraction)
'voice': NoteFeatureSpec('voice', col_index=3, vocab_size=25), # [0, 24] (20 unique)
'staff': NoteFeatureSpec('staff', col_index=13, vocab_size=37), # [0, 36]
# Pitch features
'pitch_step': NoteFeatureSpec('pitch_step', col_index=4, vocab_size=7), # [0, 6] C...B
'pitch_alter': NoteFeatureSpec('pitch_alter', col_index=5, vocab_size=5), # [0, 4] (shifted +2 in extraction)
'pitch_octave': NoteFeatureSpec('pitch_octave', col_index=6, vocab_size=15, shift=3), # [-3, 11] -> [0, 14]
# Duration (tokenised)
'duration': NoteFeatureSpec('duration', col_index=7, vocab_size=727), # [0, 726] (filtered vocab)
# Context: clef, time signature, key
'clef': NoteFeatureSpec('clef', col_index=8, vocab_size=6), # [0, 5]
'ts_beats': NoteFeatureSpec('ts_beats', col_index=9, vocab_size=57, remap=TS_BEATS_REMAP), # 57 unique values
'ts_beat_type': NoteFeatureSpec('ts_beat_type', col_index=10, vocab_size=6, remap=TS_BEAT_TYPE_REMAP), # 6 values: {1,2,4,8,16,32}
'key_fifths': NoteFeatureSpec('key_fifths', col_index=11, vocab_size=15, shift=7), # [-7, +7] -> [0, 14]
'key_mode': NoteFeatureSpec('key_mode', col_index=12, vocab_size=2), # 0=major, 1=minor (optional)
}
class ScoreGraphMultiFeatureDataset(Dataset):
"""
Dataset for loading score graphs with multiple configurable features.
Supports:
- Multiple discrete features per note
- Graph structure (edges) for potential GNN integration
- Configurable feature selection for ablation studies
Args:
graph_dir: Directory containing .pt graph files
features: List of feature names to extract (from FEATURE_SPECS)
file_list: Optional list of specific files to use
max_notes: Maximum number of notes per graph (filter out larger)
max_bars: Maximum number of bars per graph (filter out larger)
identifier_pool_size: Size of entity ID pool
include_graph: Whether to include graph structure (edges)
id_assignment: 'sequential' (0,1,2,...) or 'random' (random permutation)
seed: Random seed for entity ID assignment
"""
def __init__(
self,
graph_dir: str,
features: List[str] = ['grid_position'],
file_list: Optional[List[str]] = None,
max_notes: Optional[int] = 256,
max_bars: Optional[int] = None,
identifier_pool_size: int = 512,
include_graph: bool = False,
id_assignment: str = 'sequential',
seed: int = 42,
):
self.graph_dir = Path(graph_dir)
self.features = features
self.max_notes = max_notes
self.max_bars = max_bars
self.identifier_pool_size = identifier_pool_size
self.include_graph = include_graph
self.id_assignment = id_assignment
self.seed = seed
# Validate features
for feat in features:
if feat not in FEATURE_SPECS:
raise ValueError(f"Unknown feature: {feat}. Available: {list(FEATURE_SPECS.keys())}")
self.feature_specs = [FEATURE_SPECS[f] for f in features]
# Track if file_list was provided (for filtering logic)
self._has_file_list = file_list is not None
# Load file list
if file_list is not None:
self.graph_files = [self.graph_dir / f for f in file_list]
self.graph_files = [f for f in self.graph_files if f.exists()]
else:
self.graph_files = sorted(self.graph_dir.glob("*.pt"))
# Filter by max_notes and max_bars
if max_notes is not None or max_bars is not None:
self._filter_by_constraints()
print(f"ScoreGraphMultiFeatureDataset: {len(self.graph_files)} graphs")
print(f"\tFeatures: {features}")
print(f"\tMax notes: {max_notes}, Max bars: {max_bars}, ID pool: {identifier_pool_size}")
def _filter_by_constraints(self):
"""Filter graphs that exceed max_notes or max_bars, using cache when available."""
# Build cache filename based on constraints
cache_parts = []
if self.max_notes is not None:
cache_parts.append(f"notes{self.max_notes}")
if self.max_bars is not None:
cache_parts.append(f"bars{self.max_bars}")
cache_name = "_".join(cache_parts)
cache_file = self.graph_dir / f".filtered_files_{cache_name}.txt"
if cache_file.exists():
with open(cache_file) as f:
cached_filenames = set(f.read().strip().split('\n'))
if self._has_file_list:
# Intersect with provided file_list (don't replace!)
current_filenames = {f.name for f in self.graph_files}
valid_filenames = current_filenames & cached_filenames
self.graph_files = [f for f in self.graph_files if f.name in valid_filenames]
else:
# Use cache directly - sort for consistent ordering
self.graph_files = sorted([self.graph_dir / fn for fn in cached_filenames if (self.graph_dir / fn).exists()])
else:
# Build cache by checking each file
filtered_files = []
for f in self.graph_files:
try:
data = torch.load(f, weights_only=False)
if self._check_constraints(data):
filtered_files.append(f)
except Exception:
continue
# Save cache (only when not using file_list)
if not self._has_file_list:
with open(cache_file, 'w') as cf:
cf.write('\n'.join(f.name for f in filtered_files))
self.graph_files = filtered_files
def _check_constraints(self, data) -> bool:
"""Check if a graph satisfies max_notes and max_bars constraints."""
if not isinstance(data, dict) or 'graph' not in data:
return False
graph = data['graph']
if 'note' not in graph.node_types:
return False
note_x = graph['note'].x
num_notes = note_x.shape[0]
# Check max_notes constraint
if self.max_notes is not None and num_notes > self.max_notes:
return False
# Check max_bars constraint (measure_idx is col 2, already 0-based)
if self.max_bars is not None:
bar_idx_col = note_x[:, 2] # col_index for measure_idx
max_bar = bar_idx_col.max().item() + 1 # +1 since measure_idx is 0-based
if max_bar > self.max_bars:
return False
return True
def _get_num_notes(self, data) -> Optional[int]:
"""Extract number of notes from graph['note'].x."""
if isinstance(data, dict) and 'graph' in data:
graph = data['graph']
if 'note' in graph.node_types:
return graph['note'].x.shape[0]
# Fallback to num_notes key
if 'num_notes' in data:
return data['num_notes']
return None
def __len__(self):
return len(self.graph_files)
def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
"""Load a graph and extract configured features from graph['note'].x."""
graph_path = self.graph_files[idx]
data = torch.load(graph_path, weights_only=False)
# Get the HeteroData graph and note features
graph = data['graph']
note_x = graph['note'].x # Shape: [num_notes, 14]
num_notes = note_x.shape[0]
# Generate unique entity IDs for this sample
if self.id_assignment == 'sequential':
entity_ids = torch.arange(num_notes, dtype=torch.long)
else:
rng = np.random.RandomState(self.seed + idx)
entity_ids = torch.from_numpy(
rng.choice(self.identifier_pool_size, size=num_notes, replace=False)
).long()
# Extract features from graph['note'].x columns
result = {
'entity_ids': entity_ids,
'num_entities': num_notes,
}
for spec in self.feature_specs:
# Get values from the column index
values = note_x[:, spec.col_index].long()
# Apply remapping if specified (for sparse value sets like ts_beat_type)
if spec.remap is not None:
# Vectorized remapping using a lookup tensor.
# Values not in the remap dict get mapped to the nearest known key.
max_key = max(spec.remap.keys())
remap_tensor = torch.zeros(max_key + 1, dtype=torch.long)
for old_val, new_val in spec.remap.items():
remap_tensor[old_val] = new_val
clamped = values.clamp(0, max_key)
# For values not in the remap, clamp ensures a valid index;
# stray values map to 0 which is acceptable as fallback.
values = remap_tensor[clamped]
elif spec.shift != 0:
# Apply shift if needed (only if no remap)
values = values + spec.shift
# Clamp to valid range
values = values.clamp(0, spec.vocab_size - 1)
result[spec.name] = values
# Include full graph if requested (for HGT pre-processing)
if self.include_graph:
result['graph'] = graph
return result
def get_vocab_sizes(self) -> Dict[str, int]:
"""Get vocabulary sizes for configured features."""
return {spec.name: spec.vocab_size for spec in self.feature_specs}
def collate_fn(batch: List[Dict]) -> Dict[str, torch.Tensor]:
"""
Collate function for multi-feature dataset.
Pads all samples to the maximum number of entities in the batch.
When graphs are present (HGT mode), extracts edge_dicts for each sample.
"""
from src.model.note_hgt import NoteHGT
batch_size = len(batch)
max_entities = max(sample['num_entities'] for sample in batch)
# Get feature names from first sample (excluding special keys)
special_keys = {'entity_ids', 'num_entities', 'edge_index', 'graph'}
feature_names = [k for k in batch[0].keys() if k not in special_keys]
# Initialize tensors
entity_ids = torch.zeros(batch_size, max_entities, dtype=torch.long)
mask = torch.zeros(batch_size, max_entities, dtype=torch.bool)
num_entities = torch.tensor([s['num_entities'] for s in batch])
features = {name: torch.zeros(batch_size, max_entities, dtype=torch.long)
for name in feature_names}
# Fill tensors
for i, sample in enumerate(batch):
n = sample['num_entities']
entity_ids[i, :n] = sample['entity_ids']
mask[i, :n] = True
for name in feature_names:
features[name][i, :n] = sample[name]
result = {
'entity_ids': entity_ids,
'mask': mask,
'num_entities': num_entities,
**features,
}
# Handle graph data if present (for HGT): extract edge_dicts
if 'graph' in batch[0]:
result['edge_dicts'] = [
NoteHGT.extract_edge_dict(s['graph']) for s in batch
]
return result
|