File size: 43,274 Bytes
d9c7e7e 3e0d8bd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 |
"""
# Melody Generation Model Development
# Project: Opentunes.ai
This notebook implements a Transformer-based melody generation model.
The model takes text prompts and generates musical melodies in MIDI format.
Key Features:
- Text-to-melody generation
- MIDI file handling
- Transformer architecture
- Training pipeline integration with HuggingFace
Note: This is a starting point and might need adjustments based on:
- Specific musical requirements
- Available training data
- Computational resources
- Desired output format
"""
import torch
import torch.nn as nn
from transformers import (
AutoModelForAudio,
AutoTokenizer,
Trainer,
TrainingArguments
)
import librosa
import numpy as np
import pandas as pd
import music21
from pathlib import Path
import json
import wandb # for experiment tracking
# =====================================
# 1. Data Loading and Preprocessing
# =====================================
class MelodyDataset(torch.utils.data.Dataset):
"""
Custom Dataset class for handling melody data.
This class:
- Loads MIDI files from a directory
- Converts MIDI files to sequences of notes and durations
- Provides data in format suitable for model training
Args:
data_dir (str): Directory containing MIDI files
max_length (int): Maximum sequence length (default: 512)
Features:
- Handles variable-length MIDI files
- Converts complex MIDI structures to simple note sequences
- Implements efficient data loading and preprocessing
"""
def __init__(self, data_dir, max_length=512):
self.data_dir = Path(data_dir)
self.max_length = max_length
self.midi_files = list(self.data_dir.glob("*.mid"))
# Initialize tokenizer for text prompts
self.tokenizer = AutoTokenizer.from_pretrained("t5-small")
print(f"Found {len(self.midi_files)} MIDI files in {data_dir}")
def midi_to_sequence(self, midi_path):
"""
Convert MIDI file to sequence of notes.
Args:
midi_path (Path): Path to MIDI file
Returns:
list: List of dictionaries containing note information
Each dict has 'pitch', 'duration', and 'offset'
Example output:
[
{'pitch': 60, 'duration': 1.0, 'offset': 0.0}, # Middle C, quarter note
{'pitch': 64, 'duration': 0.5, 'offset': 1.0}, # E, eighth note
...
]
"""
score = music21.converter.parse(str(midi_path))
notes = []
# Extract notes and their properties
for n in score.flat.notesAndRests:
if isinstance(n, music21.note.Note):
notes.append({
'pitch': n.pitch.midi, # MIDI pitch number (0-127)
'duration': n.duration.quarterLength, # Duration in quarter notes
'offset': n.offset # Start time in quarter notes
})
return notes
def __getitem__(self, idx):
"""
Get a single item from the dataset.
Args:
idx (int): Index of the item
Returns:
dict: Dictionary containing:
- 'notes': Tensor of note pitches
- 'durations': Tensor of note durations
Note: Both tensors are padded/truncated to max_length
"""
midi_file = self.midi_files[idx]
melody_sequence = self.midi_to_sequence(midi_file)
# Convert to tensors with padding/truncation
notes = torch.tensor([n['pitch'] for n in melody_sequence])
durations = torch.tensor([n['duration'] for n in melody_sequence])
# Pad or truncate sequences
if len(notes) < self.max_length:
# Pad with rest values
pad_length = self.max_length - len(notes)
notes = torch.cat([notes, torch.zeros(pad_length)])
durations = torch.cat([durations, torch.zeros(pad_length)])
else:
# Truncate to max_length
notes = notes[:self.max_length]
durations = durations[:self.max_length]
return {
'notes': notes,
'durations': durations,
}
def __len__(self):
return len(self.midi_files)
# =====================================
# 2. Model Architecture Development
# =====================================
class MelodyTransformer(nn.Module):
"""
Transformer-based model for melody generation.
Architecture Overview:
1. Embedding layers for notes, durations, and positions
2. Transformer encoder for sequence processing
3. Separate prediction heads for notes and durations
Args:
num_notes (int): Size of note vocabulary (default: 128 for MIDI range)
max_duration (int): Number of possible duration values (default: 32)
d_model (int): Dimension of the model (default: 512)
nhead (int): Number of attention heads (default: 8)
num_layers (int): Number of transformer layers (default: 6)
Forward Pass:
- Input: note sequence, duration sequence, position indices
- Output: predictions for next note and duration
"""
def __init__(self,
num_notes=128, # MIDI note range (0-127)
max_duration=32, # Quantized duration values
d_model=512, # Model dimension (as in original Transformer)
nhead=8, # Multi-head attention
num_layers=6): # Number of Transformer layers
super().__init__()
# Embedding layers
self.note_embedding = nn.Embedding(
num_embeddings=num_notes,
embedding_dim=d_model,
padding_idx=0 # Use 0 for padding
)
self.duration_embedding = nn.Embedding(
num_embeddings=max_duration,
embedding_dim=d_model,
padding_idx=0
)
self.position_embedding = nn.Embedding(
num_embeddings=1024, # Maximum sequence length
embedding_dim=d_model
)
# Transformer architecture
encoder_layer = nn.TransformerEncoderLayer(
d_model=d_model,
nhead=nhead,
dim_feedforward=4*d_model, # As per original Transformer paper
dropout=0.1,
activation='gelu' # Modern activation function
)
self.transformer = nn.TransformerEncoder(
encoder_layer=encoder_layer,
num_layers=num_layers,
norm=nn.LayerNorm(d_model)
)
# Output heads
self.note_head = nn.Sequential(
nn.Linear(d_model, d_model),
nn.ReLU(),
nn.Dropout(0.1),
nn.Linear(d_model, num_notes)
)
self.duration_head = nn.Sequential(
nn.Linear(d_model, d_model),
nn.ReLU(),
nn.Dropout(0.1),
nn.Linear(d_model, max_duration)
)
def forward(self, notes, durations, positions):
"""
Forward pass through the model.
Args:
notes (torch.Tensor): Shape [batch_size, seq_length]
Contains MIDI note numbers
durations (torch.Tensor): Shape [batch_size, seq_length]
Contains quantized duration values
positions (torch.Tensor): Shape [batch_size, seq_length]
Contains position indices
Returns:
tuple: (note_logits, duration_logits)
- note_logits: Shape [batch_size, seq_length, num_notes]
- duration_logits: Shape [batch_size, seq_length, max_duration]
Note:
The model predicts both the next note and its duration
simultaneously, allowing for coherent melody generation.
"""
# Get embeddings for each component
note_emb = self.note_embedding(notes) # [B, S, D]
duration_emb = self.duration_embedding(durations) # [B, S, D]
pos_emb = self.position_embedding(positions) # [B, S, D]
# Combine embeddings
# Sum embeddings as in original Transformer paper
x = note_emb + duration_emb + pos_emb # [B, S, D]
# Apply Transformer
# Note: Need to reshape for Transformer which expects [S, B, D]
x = x.transpose(0, 1)
x = self.transformer(x)
x = x.transpose(0, 1) # Back to [B, S, D]
# Generate predictions
note_logits = self.note_head(x) # [B, S, num_notes]
duration_logits = self.duration_head(x) # [B, S, max_duration]
return note_logits, duration_logits
def generate(self, prompt, max_length=512, temperature=1.0):
"""
Generate a melody from a starting prompt.
Args:
prompt (dict): Initial notes and durations
max_length (int): Maximum sequence length to generate
temperature (float): Sampling temperature (higher = more random)
Returns:
tuple: (generated_notes, generated_durations)
Example:
>>> model = MelodyTransformer()
>>> prompt = {'notes': [60, 64, 67], 'durations': [1.0, 1.0, 1.0]}
>>> notes, durations = model.generate(prompt)
"""
self.eval() # Set to evaluation mode
with torch.no_grad():
# Initialize with prompt
current_notes = torch.tensor(prompt['notes']).unsqueeze(0)
current_durations = torch.tensor(prompt['durations']).unsqueeze(0)
generated_notes = list(prompt['notes'])
generated_durations = list(prompt['durations'])
# Generate one note at a time
for i in range(len(prompt['notes']), max_length):
# Create position tensor
positions = torch.arange(len(generated_notes)).unsqueeze(0)
# Get predictions
note_logits, duration_logits = self(
current_notes,
current_durations,
positions
)
# Sample from logits using temperature
note_probs = F.softmax(note_logits[:, -1] / temperature, dim=-1)
duration_probs = F.softmax(duration_logits[:, -1] / temperature, dim=-1)
next_note = torch.multinomial(note_probs, 1)
next_duration = torch.multinomial(duration_probs, 1)
# Append to generated sequence
generated_notes.append(next_note.item())
generated_durations.append(next_duration.item())
# Update current sequence
current_notes = torch.tensor(generated_notes).unsqueeze(0)
current_durations = torch.tensor(generated_durations).unsqueeze(0)
return generated_notes, generated_durations
# =====================================
# 3. Training Pipeline
# =====================================
class MelodyTrainer:
"""
Custom training pipeline for the melody generation model.
Features:
- Automated training loop
- Validation monitoring
- Checkpoint saving
- Logging and metrics tracking
Args:
model (MelodyTransformer): The model to train
config (dict): Training configuration
device (str): Device to train on ('cuda' or 'cpu')
"""
def __init__(self, model, config, device='cuda'):
self.model = model.to(device)
self.config = config
self.device = device
# Initialize training components
self.criterion = nn.CrossEntropyLoss(ignore_index=0) # Ignore padding
self.optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=config['learning_rate'],
weight_decay=config.get('weight_decay', 0.01)
)
# Learning rate scheduler
self.scheduler = torch.optim.lr_scheduler.OneCycleLR(
self.optimizer,
max_lr=config['learning_rate'],
epochs=config['epochs'],
steps_per_epoch=config['steps_per_epoch']
)
# Initialize wandb for experiment tracking
if config.get('use_wandb', False):
wandb.init(
project="opentunes-melody",
config=config,
name=f"melody_training_{datetime.now().strftime('%Y%m%d_%H%M')}"
)
def train_epoch(self, train_loader):
"""
Train for one epoch.
Args:
train_loader (DataLoader): Training data loader
Returns:
dict: Training metrics for this epoch
"""
self.model.train()
epoch_loss = 0
epoch_note_acc = 0
epoch_dur_acc = 0
num_batches = 0
for batch in tqdm(train_loader, desc="Training"):
# Move batch to device
notes = batch['notes'].to(self.device)
durations = batch['durations'].to(self.device)
positions = torch.arange(notes.size(1)).unsqueeze(0).expand(
notes.size(0), -1).to(self.device)
# Forward pass
note_logits, duration_logits = self.model(notes, durations, positions)
# Calculate loss
# Shift sequences for next-token prediction
note_loss = self.criterion(
note_logits[:, :-1].reshape(-1, note_logits.size(-1)),
notes[:, 1:].reshape(-1)
)
duration_loss = self.criterion(
duration_logits[:, :-1].reshape(-1, duration_logits.size(-1)),
durations[:, 1:].reshape(-1)
)
loss = note_loss + duration_loss
# Backward pass
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
self.scheduler.step()
# Calculate metrics
with torch.no_grad():
note_preds = note_logits.argmax(dim=-1)
dur_preds = duration_logits.argmax(dim=-1)
note_acc = (note_preds[:, :-1] == notes[:, 1:]).float().mean()
dur_acc = (dur_preds[:, :-1] == durations[:, 1:]).float().mean()
# Update running metrics
epoch_loss += loss.item()
epoch_note_acc += note_acc.item()
epoch_dur_acc += dur_acc.item()
num_batches += 1
# Log batch metrics
if self.config.get('use_wandb', False):
wandb.log({
'batch_loss': loss.item(),
'note_accuracy': note_acc.item(),
'duration_accuracy': dur_acc.item(),
'learning_rate': self.scheduler.get_last_lr()[0]
})
# Calculate epoch metrics
metrics = {
'loss': epoch_loss / num_batches,
'note_accuracy': epoch_note_acc / num_batches,
'duration_accuracy': epoch_dur_acc / num_batches
}
return metrics
def validate(self, val_loader):
"""
Validate the model.
Args:
val_loader (DataLoader): Validation data loader
Returns:
dict: Validation metrics
"""
self.model.eval()
val_loss = 0
val_note_acc = 0
val_dur_acc = 0
num_batches = 0
with torch.no_grad():
for batch in tqdm(val_loader, desc="Validation"):
notes = batch['notes'].to(self.device)
durations = batch['durations'].to(self.device)
positions = torch.arange(notes.size(1)).unsqueeze(0).expand(
notes.size(0), -1).to(self.device)
# Forward pass
note_logits, duration_logits = self.model(notes, durations, positions)
# Calculate metrics (similar to training)
note_loss = self.criterion(
note_logits[:, :-1].reshape(-1, note_logits.size(-1)),
notes[:, 1:].reshape(-1)
)
duration_loss = self.criterion(
duration_logits[:, :-1].reshape(-1, duration_logits.size(-1)),
durations[:, 1:].reshape(-1)
)
loss = note_loss + duration_loss
note_preds = note_logits.argmax(dim=-1)
dur_preds = duration_logits.argmax(dim=-1)
note_acc = (note_preds[:, :-1] == notes[:, 1:]).float().mean()
dur_acc = (dur_preds[:, :-1] == durations[:, 1:]).float().mean()
val_loss += loss.item()
val_note_acc += note_acc.item()
val_dur_acc += dur_acc.item()
num_batches += 1
metrics = {
'val_loss': val_loss / num_batches,
'val_note_accuracy': val_note_acc / num_batches,
'val_duration_accuracy': val_dur_acc / num_batches
}
return metrics
def train(self, train_loader, val_loader):
"""
Full training loop.
Args:
train_loader (DataLoader): Training data loader
val_loader (DataLoader): Validation data loader
"""
best_val_loss = float('inf')
for epoch in range(self.config['epochs']):
print(f"\nEpoch {epoch+1}/{self.config['epochs']}")
# Training phase
train_metrics = self.train_epoch(train_loader)
print(f"Training metrics: {train_metrics}")
# Validation phase
val_metrics = self.validate(val_loader)
print(f"Validation metrics: {val_metrics}")
# Save checkpoint if best so far
if val_metrics['val_loss'] < best_val_loss:
best_val_loss = val_metrics['val_loss']
self.save_checkpoint(
f"models/melody-gen/weights/v0.1.0/best_model.pth",
epoch,
train_metrics,
val_metrics
)
# Log epoch metrics
if self.config.get('use_wandb', False):
wandb.log({
'epoch': epoch,
**train_metrics,
**val_metrics
})
def save_checkpoint(self, path, epoch, train_metrics, val_metrics):
"""
Save model checkpoint.
Args:
path (str): Path to save checkpoint
epoch (int): Current epoch
train_metrics (dict): Training metrics
val_metrics (dict): Validation metrics
"""
checkpoint = {
'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'train_metrics': train_metrics,
'val_metrics': val_metrics,
'config': self.config
}
torch.save(checkpoint, path)
print(f"Checkpoint saved to {path}")
# =====================================
# 4. Evaluation Functions
# =====================================
class MelodyEvaluator:
"""
Comprehensive evaluation suite for melody generation models.
Features:
- Note accuracy metrics
- Musical quality assessment
- Style consistency checking
- Sample generation and analysis
Args:
model (MelodyTransformer): Trained model to evaluate
device (str): Device to run evaluation on
"""
def __init__(self, model, device='cuda'):
self.model = model.to(device)
self.device = device
self.model.eval() # Set model to evaluation mode
def evaluate_metrics(self, test_loader):
"""
Compute quantitative metrics on test set.
Args:
test_loader (DataLoader): Test data loader
Returns:
dict: Dictionary of evaluation metrics
"""
metrics = {
'note_accuracy': 0,
'rhythm_accuracy': 0,
'sequence_coherence': 0,
'scale_consistency': 0
}
num_batches = 0
with torch.no_grad():
for batch in tqdm(test_loader, desc="Evaluating"):
notes = batch['notes'].to(self.device)
durations = batch['durations'].to(self.device)
positions = torch.arange(notes.size(1)).unsqueeze(0).expand(
notes.size(0), -1).to(self.device)
# Get model predictions
note_logits, duration_logits = self.model(notes, durations, positions)
# Calculate basic accuracy
note_preds = note_logits.argmax(dim=-1)
dur_preds = duration_logits.argmax(dim=-1)
metrics['note_accuracy'] += (note_preds[:, :-1] == notes[:, 1:]).float().mean().item()
metrics['rhythm_accuracy'] += (dur_preds[:, :-1] == durations[:, 1:]).float().mean().item()
# Calculate musical coherence metrics
metrics['sequence_coherence'] += self._calculate_coherence(note_preds)
metrics['scale_consistency'] += self._check_scale_consistency(note_preds)
num_batches += 1
# Average metrics
for key in metrics:
metrics[key] /= num_batches
return metrics
def _calculate_coherence(self, note_sequence):
"""
Calculate musical coherence score.
Checks for:
- Melodic intervals (steps vs leaps)
- Phrase structure
- Repetition patterns
Args:
note_sequence (torch.Tensor): Predicted note sequence
Returns:
float: Coherence score between 0 and 1
"""
# Convert to numpy for music21 processing
notes = note_sequence.cpu().numpy()
# Calculate interval distribution
intervals = np.diff(notes, axis=1)
step_ratio = np.mean(np.abs(intervals) <= 2) # Proportion of stepwise motion
# Check for phrase repetition
phrase_score = self._analyze_phrases(notes)
# Combine metrics
coherence_score = 0.6 * step_ratio + 0.4 * phrase_score
return coherence_score
def _check_scale_consistency(self, note_sequence):
"""
Check if generated notes follow consistent scale patterns.
Args:
note_sequence (torch.Tensor): Predicted note sequence
Returns:
float: Scale consistency score between 0 and 1
"""
notes = note_sequence.cpu().numpy()
# Create pitch class histogram
pitch_classes = notes % 12
histogram = np.bincount(pitch_classes.flatten(), minlength=12)
# Check against common scales
major_scale = np.array([1, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1])
minor_scale = np.array([1, 0, 1, 1, 0, 1, 0, 1, 1, 0, 1, 0])
# Calculate consistency scores
major_score = np.sum((histogram > 0) == major_scale) / 12
minor_score = np.sum((histogram > 0) == minor_scale) / 12
return max(major_score, minor_score)
def generate_and_evaluate_samples(self, num_samples=10, max_length=512):
"""
Generate and evaluate multiple melody samples.
Args:
num_samples (int): Number of samples to generate
max_length (int): Maximum length of each sample
Returns:
tuple: (generated_samples, evaluation_results)
"""
samples = []
results = []
for i in range(num_samples):
# Generate sample
prompt = {
'notes': [60], # Start with middle C
'durations': [1.0] # Quarter note
}
notes, durations = self.model.generate(
prompt,
max_length=max_length,
temperature=0.8
)
# Evaluate sample
sample_metrics = {
'melodic_range': self._calculate_melodic_range(notes),
'rhythm_variety': self._calculate_rhythm_variety(durations),
'musical_coherence': self._evaluate_musical_qualities(notes, durations)
}
samples.append({'notes': notes, 'durations': durations})
results.append(sample_metrics)
# Save generated sample
self._save_sample(
notes,
durations,
f"models/melody-gen/examples/generated_samples/sample_{i+1}.mid"
)
return samples, results
def _calculate_melodic_range(self, notes):
"""
Calculate the melodic range and distribution.
Args:
notes (list): List of MIDI note numbers
Returns:
dict: Melodic range statistics
"""
return {
'range': max(notes) - min(notes),
'mean': np.mean(notes),
'std': np.std(notes)
}
def _calculate_rhythm_variety(self, durations):
"""
Analyze rhythm patterns and variety.
Args:
durations (list): List of note durations
Returns:
dict: Rhythm statistics
"""
return {
'unique_values': len(set(durations)),
'variance': np.var(durations),
'pattern_complexity': len(set(zip(durations[:-1], durations[1:])))
}
def _evaluate_musical_qualities(self, notes, durations):
"""
Evaluate musical qualities of the generated melody.
Checks for:
- Phrase structure
- Melodic contour
- Rhythmic patterns
- Musical tension and resolution
Args:
notes (list): List of MIDI note numbers
durations (list): List of note durations
Returns:
dict: Musical quality metrics
"""
# Convert to music21 stream for analysis
stream = self._create_music21_stream(notes, durations)
return {
'phrase_structure': self._analyze_phrases(stream),
'melodic_contour': self._analyze_contour(notes),
'rhythmic_complexity': self._analyze_rhythm(durations),
'tension_resolution': self._analyze_tension(notes)
}
def _save_sample(self, notes, durations, filepath):
"""
Save generated sample as MIDI file.
Args:
notes (list): List of MIDI note numbers
durations (list): List of note durations
filepath (str): Path to save MIDI file
"""
stream = music21.stream.Stream()
for note, duration in zip(notes, durations):
n = music21.note.Note(note)
n.duration = music21.duration.Duration(duration)
stream.append(n)
stream.write('midi', fp=filepath)
def generate_evaluation_report(self, test_loader):
"""
Generate comprehensive evaluation report.
Args:
test_loader (DataLoader): Test data loader
Returns:
dict: Complete evaluation report
"""
# Basic metrics
metrics = self.evaluate_metrics(test_loader)
# Generate and evaluate samples
samples, sample_results = self.generate_and_evaluate_samples()
# Compile complete report
report = {
'quantitative_metrics': metrics,
'sample_evaluations': sample_results,
'generation_timestamp': datetime.now().strftime('%Y-%m-%d %H:%M:%S'),
'model_version': '0.1.0'
}
# Save report
with open('models/melody-gen/examples/evaluation_report.json', 'w') as f:
json.dump(report, f, indent=2)
return report
# =====================================
# 5. Generation and Inference
# =====================================
class MelodyGenerator:
"""
High-level interface for generating melodies using trained model.
Features:
- Text-to-melody generation
- Style conditioning
- Batch generation
- Format conversion and export
Args:
model (MelodyTransformer): Trained model
device (str): Device to run generation on
config (dict): Generation parameters
"""
def __init__(self, model, device='cuda', config=None):
self.model = model.to(device)
self.device = device
self.model.eval()
# Default generation config
self.config = {
'temperature': 0.8,
'max_length': 512,
'top_k': 50,
'top_p': 0.95,
'repetition_penalty': 1.2
}
if config:
self.config.update(config)
def generate_from_prompt(self, prompt, style=None):
"""
Generate melody from text prompt.
Args:
prompt (str): Text description of desired melody
style (dict, optional): Style parameters
{
'genre': 'pop/jazz/classical',
'tempo': beats per minute,
'mood': 'happy/sad/energetic'
}
Returns:
dict: Generated melody information
{
'notes': List of MIDI notes,
'durations': List of note durations,
'midi_path': Path to saved MIDI file,
'metadata': Generation metadata
}
"""
# Process prompt and style
generation_params = self._prepare_generation_params(prompt, style)
with torch.no_grad():
# Initialize sequence with start token
current_notes = torch.tensor([[60]]).to(self.device) # Middle C
current_durations = torch.tensor([[1.0]]).to(self.device) # Quarter note
generated_notes = []
generated_durations = []
# Generate sequence
for i in range(self.config['max_length']):
# Get position encoding
position = torch.arange(current_notes.size(1)).unsqueeze(0).to(self.device)
# Get predictions
note_logits, duration_logits = self.model(
current_notes,
current_durations,
position
)
# Apply temperature and sampling strategies
next_note = self._sample_from_logits(
note_logits[:, -1],
temperature=generation_params['temperature'],
top_k=generation_params['top_k'],
top_p=generation_params['top_p']
)
next_duration = self._sample_from_logits(
duration_logits[:, -1],
temperature=generation_params['temperature']
)
# Apply repetition penalty
if len(generated_notes) > 0:
next_note = self._apply_repetition_penalty(
next_note,
generated_notes,
generation_params['repetition_penalty']
)
# Append to sequences
generated_notes.append(next_note.item())
generated_durations.append(next_duration.item())
# Update input sequences
current_notes = torch.tensor([generated_notes]).to(self.device)
current_durations = torch.tensor([generated_durations]).to(self.device)
# Check for end condition
if self._check_end_condition(generated_notes, generated_durations):
break
# Post-process and save
return self._post_process_and_save(
generated_notes,
generated_durations,
prompt,
style
)
def batch_generate(self, prompts, styles=None):
"""
Generate multiple melodies in batch.
Args:
prompts (list): List of text prompts
styles (list, optional): List of style parameters
Returns:
list: List of generated melodies
"""
results = []
for i, prompt in enumerate(prompts):
style = styles[i] if styles else None
result = self.generate_from_prompt(prompt, style)
results.append(result)
return results
def _prepare_generation_params(self, prompt, style):
"""
Prepare generation parameters based on prompt and style.
Args:
prompt (str): Text prompt
style (dict): Style parameters
Returns:
dict: Generation parameters
"""
params = self.config.copy()
# Adjust parameters based on style
if style:
if style.get('genre') == 'classical':
params['temperature'] *= 0.9 # More conservative
params['repetition_penalty'] *= 1.1
elif style.get('genre') == 'jazz':
params['temperature'] *= 1.1 # More experimental
params['top_k'] *= 1.2
if style.get('mood') == 'energetic':
params['temperature'] *= 1.1
elif style.get('mood') == 'calm':
params['temperature'] *= 0.9
return params
def _sample_from_logits(self, logits, temperature=1.0, top_k=None, top_p=None):
"""
Sample from logits with temperature and optional top-k/top-p filtering.
Args:
logits (torch.Tensor): Raw logits
temperature (float): Sampling temperature
top_k (int, optional): Top-k filtering parameter
top_p (float, optional): Nucleus sampling parameter
Returns:
torch.Tensor: Sampled token
"""
logits = logits / temperature
# Top-k filtering
if top_k is not None:
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits[indices_to_remove] = float('-inf')
# Top-p filtering (nucleus sampling)
if top_p is not None:
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
indices_to_remove = sorted_indices_to_remove.scatter(
dim=-1,
index=sorted_indices,
src=sorted_indices_to_remove
)
logits[indices_to_remove] = float('-inf')
# Sample
probs = F.softmax(logits, dim=-1)
return torch.multinomial(probs, 1)
def _post_process_and_save(self, notes, durations, prompt, style):
"""
Post-process and save generated melody.
Args:
notes (list): Generated notes
durations (list): Generated durations
prompt (str): Original prompt
style (dict): Style parameters
Returns:
dict: Generation results and metadata
"""
# Create timestamp
timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
# Create MIDI file
midi_path = f"models/melody-gen/examples/generated_samples/melody_{timestamp}.mid"
self._save_to_midi(notes, durations, midi_path)
# Prepare metadata
metadata = {
'timestamp': timestamp,
'prompt': prompt,
'style': style,
'generation_params': self.config,
'stats': {
'length': len(notes),
'pitch_range': max(notes) - min(notes),
'unique_pitches': len(set(notes)),
'unique_durations': len(set(durations))
}
}
# Save metadata
metadata_path = f"models/melody-gen/examples/generated_samples/melody_{timestamp}.json"
with open(metadata_path, 'w') as f:
json.dump(metadata, f, indent=2)
return {
'notes': notes,
'durations': durations,
'midi_path': midi_path,
'metadata': metadata
}
# =====================================
# 6. Utility Functions and Helpers
# =====================================
class MelodyUtils:
"""
Utility functions for melody processing and manipulation.
"""
@staticmethod
def save_to_midi(notes, durations, path):
"""
Save melody to MIDI file with enhanced musical properties.
Args:
notes (list): MIDI note numbers
durations (list): Note durations
path (str): Output path
"""
stream = music21.stream.Stream()
# Add time signature and tempo
stream.append(music21.meter.TimeSignature('4/4'))
stream.append(music21.tempo.MetronomeMark(number=120))
# Add notes with velocity for dynamics
for note, duration in zip(notes, durations):
n = music21.note.Note(note)
n.duration = music21.duration.Duration(duration)
# Add velocity (dynamics) based on position in phrase
n.volume.velocity = MelodyUtils._calculate_velocity(note, notes)
stream.append(n)
stream.write('midi', fp=path)
@staticmethod
def _calculate_velocity(note, notes_sequence):
"""Calculate appropriate velocity for musical expression."""
base_velocity = 64
# Emphasize phrase beginnings and high points
if note == max(notes_sequence):
return min(base_velocity + 32, 127)
return base_velocity
# =====================================
# 7. Enhanced Generation Features
# =====================================
class EnhancedMelodyGenerator(MelodyGenerator):
"""
Extended melody generator with additional features.
"""
def generate_with_structure(self, prompt, form="AABA"):
"""
Generate melody with specific musical form.
Args:
prompt (str): Text prompt
form (str): Musical form (e.g., "AABA", "ABAC")
Returns:
dict: Generated melody with structural sections
"""
sections = {}
full_melody = []
for section in form:
if section not in sections:
# Generate new section
result = self.generate_from_prompt(
prompt + f" for section {section}",
{'section': section}
)
sections[section] = (result['notes'], result['durations'])
# Add section to full melody
notes, durations = sections[section]
full_melody.extend(zip(notes, durations))
return self._post_process_structured_melody(full_melody, form)
def generate_with_harmony(self, prompt, chord_progression=None):
"""
Generate melody with harmonic constraints.
Args:
prompt (str): Text prompt
chord_progression (list): Optional chord progression
Returns:
dict: Generated melody with harmonic context
"""
if chord_progression is None:
chord_progression = self._generate_chord_progression()
# Generate melody considering harmony
generation_params = self._prepare_generation_params(prompt, {
'harmony': chord_progression
})
return self.generate_from_prompt(prompt, generation_params)
# =====================================
# 8. Example Usage Scenarios
# =====================================
def example_usage():
"""Example usage of the melody generation system."""
# 1. Basic melody generation
generator = MelodyGenerator(model)
result = generator.generate_from_prompt(
"Create an upbeat pop melody in C major"
)
# 2. Style-conditional generation
styled_result = generator.generate_from_prompt(
"Create a jazz melody",
style={
'genre': 'jazz',
'tempo': 120,
'mood': 'energetic'
}
)
# 3. Structured generation
enhanced_generator = EnhancedMelodyGenerator(model)
structured_result = enhanced_generator.generate_with_structure(
"Create a memorable melody",
form="AABA"
)
# 4. Batch generation
prompts = [
"Happy birthday song style",
"Sad emotional melody",
"Energetic dance tune"
]
batch_results = generator.batch_generate(prompts)
# 5. Generation with harmony
harmonic_result = enhanced_generator.generate_with_harmony(
"Create a melody",
chord_progression=["C", "Am", "F", "G"]
)
return {
'basic': result,
'styled': styled_result,
'structured': structured_result,
'batch': batch_results,
'harmonic': harmonic_result
}
# =====================================
# 9. Integration Example
# =====================================
def run_complete_pipeline():
"""
Complete pipeline from training to generation.
"""
# 1. Load configuration
with open('models/melody-gen/config/model_config.json') as f:
model_config = json.load(f)
# 2. Initialize model
model = MelodyTransformer(**model_config)
# 3. Load dataset
train_dataset = MelodyDataset('datasets/train')
val_dataset = MelodyDataset('datasets/val')
test_dataset = MelodyDataset('datasets/test')
# 4. Training
trainer = MelodyTrainer(model, model_config)
trainer.train(train_dataset, val_dataset)
# 5. Evaluation
evaluator = MelodyEvaluator(model)
eval_results = evaluator.generate_evaluation_report(test_dataset)
# 6. Generation
generator = MelodyGenerator(model)
samples = generator.generate_from_prompt(
"Create an original melody",
style={'genre': 'pop', 'mood': 'happy'}
)
return {
'evaluation': eval_results,
'samples': samples
}
if __name__ == "__main__":
# Run example usage
results = example_usage()
# Run complete pipeline
pipeline_results = run_complete_pipeline() |