Spaces:
Running
on
Zero
Running
on
Zero
File size: 26,884 Bytes
358bb13 c47c661 358bb13 92cdb9d 358bb13 7c2cd8e 92cdb9d 358bb13 990bc24 358bb13 990bc24 358bb13 990bc24 358bb13 8993257 f65e448 8993257 f65e448 8993257 f65e448 8993257 f65e448 8993257 f65e448 8993257 f65e448 8993257 |
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 |
"""
LoRA Training Service
Handles fine-tuning of DiffRhythm2 model using LoRA adapters for vocal and symbolic music.
"""
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from pathlib import Path
import json
import logging
from typing import Dict, List, Optional, Callable
import soundfile as sf
import numpy as np
import time
import shutil
import zipfile
from datetime import datetime
logger = logging.getLogger(__name__)
class TrainingDataset(Dataset):
"""Dataset for LoRA training"""
def __init__(
self,
audio_files: List[str],
metadata_list: List[Dict],
sample_rate: int = 44100,
clip_length: float = 10.0
):
"""
Initialize training dataset
Args:
audio_files: List of paths to audio files
metadata_list: List of metadata dicts for each audio file
sample_rate: Target sample rate
clip_length: Length of training clips in seconds
"""
self.audio_files = audio_files
self.metadata_list = metadata_list
self.sample_rate = sample_rate
self.clip_length = clip_length
self.clip_samples = int(clip_length * sample_rate)
logger.info(f"Initialized dataset with {len(audio_files)} audio files")
def __len__(self):
return len(self.audio_files)
def __getitem__(self, idx):
"""Get training sample"""
try:
audio_path = self.audio_files[idx]
metadata = self.metadata_list[idx]
# Load audio
y, sr = sf.read(audio_path)
# Resample if needed
if sr != self.sample_rate:
import librosa
y = librosa.resample(y, orig_sr=sr, target_sr=self.sample_rate)
# Ensure mono
if y.ndim > 1:
y = y.mean(axis=1)
# Extract/pad to clip length
if len(y) > self.clip_samples:
# Random crop
start = np.random.randint(0, len(y) - self.clip_samples)
y = y[start:start + self.clip_samples]
else:
# Pad
y = np.pad(y, (0, self.clip_samples - len(y)))
# Generate prompt from metadata
prompt = self._generate_prompt(metadata)
return {
'audio': torch.FloatTensor(y),
'prompt': prompt,
'metadata': metadata
}
except Exception as e:
logger.error(f"Error loading sample {idx}: {str(e)}")
# Return empty sample on error
return {
'audio': torch.zeros(self.clip_samples),
'prompt': "",
'metadata': {}
}
def _generate_prompt(self, metadata: Dict) -> str:
"""Generate text prompt from metadata"""
parts = []
if 'genre' in metadata and metadata['genre'] != 'unknown':
parts.append(metadata['genre'])
if 'instrumentation' in metadata:
parts.append(f"with {metadata['instrumentation']}")
if 'bpm' in metadata:
parts.append(f"at {metadata['bpm']} BPM")
if 'key' in metadata:
parts.append(f"in {metadata['key']}")
if 'mood' in metadata:
parts.append(f"{metadata['mood']} mood")
if 'description' in metadata:
parts.append(metadata['description'])
return " ".join(parts) if parts else "music"
class LoRATrainingService:
"""Service for training LoRA adapters for DiffRhythm2"""
def __init__(self):
"""Initialize LoRA training service"""
self.models_dir = Path("models")
self.lora_dir = self.models_dir / "loras"
self.lora_dir.mkdir(parents=True, exist_ok=True)
self.training_data_dir = Path("training_data")
self.training_data_dir.mkdir(parents=True, exist_ok=True)
self.device = "cuda" if torch.cuda.is_available() else "cpu"
# Training state
self.is_training = False
self.current_epoch = 0
self.current_step = 0
self.training_loss = []
self.training_config = None
logger.info(f"LoRATrainingService initialized on {self.device}")
def prepare_dataset(
self,
dataset_name: str,
audio_files: List[str],
metadata_list: List[Dict],
split_ratio: float = 0.9
) -> Dict:
"""
Prepare and save training dataset
Args:
dataset_name: Name for this dataset
audio_files: List of audio file paths
metadata_list: List of metadata for each file
split_ratio: Train/validation split ratio
Returns:
Dataset information dictionary
"""
try:
logger.info(f"Preparing dataset: {dataset_name}")
# Create dataset directory
dataset_dir = self.training_data_dir / dataset_name
dataset_dir.mkdir(parents=True, exist_ok=True)
# Split into train/val
num_samples = len(audio_files)
num_train = int(num_samples * split_ratio)
indices = np.random.permutation(num_samples)
train_indices = indices[:num_train]
val_indices = indices[num_train:]
# Save metadata
dataset_info = {
'name': dataset_name,
'created': datetime.now().isoformat(),
'num_samples': num_samples,
'num_train': num_train,
'num_val': num_samples - num_train,
'train_files': [audio_files[i] for i in train_indices],
'train_metadata': [metadata_list[i] for i in train_indices],
'val_files': [audio_files[i] for i in val_indices],
'val_metadata': [metadata_list[i] for i in val_indices]
}
# Save to disk
metadata_path = dataset_dir / "dataset_info.json"
with open(metadata_path, 'w') as f:
json.dump(dataset_info, f, indent=2)
logger.info(f"Dataset prepared: {num_train} train, {num_samples - num_train} val samples")
return dataset_info
except Exception as e:
logger.error(f"Dataset preparation failed: {str(e)}")
raise
def load_dataset(self, dataset_name: str) -> Optional[Dict]:
"""Load prepared dataset information"""
try:
dataset_dir = self.training_data_dir / dataset_name
metadata_path = dataset_dir / "dataset_info.json"
if not metadata_path.exists():
logger.warning(f"Dataset not found: {dataset_name}")
return None
with open(metadata_path, 'r') as f:
return json.load(f)
except Exception as e:
logger.error(f"Failed to load dataset {dataset_name}: {str(e)}")
return None
def list_datasets(self) -> List[str]:
"""List available prepared datasets"""
try:
datasets = []
for dataset_dir in self.training_data_dir.iterdir():
if dataset_dir.is_dir() and (dataset_dir / "dataset_info.json").exists():
datasets.append(dataset_dir.name)
return datasets
except Exception as e:
logger.error(f"Failed to list datasets: {str(e)}")
return []
def list_loras(self) -> List[str]:
"""List available LoRA adapters"""
try:
loras = []
if not self.lora_dir.exists():
return loras
for lora_path in self.lora_dir.iterdir():
if lora_path.is_dir():
# Check for adapter files
if (lora_path / "adapter_config.json").exists():
loras.append(lora_path.name)
# Also check for .safetensors or .bin files
elif list(lora_path.glob("*.safetensors")) or list(lora_path.glob("*.bin")):
loras.append(lora_path.name)
return sorted(loras)
except Exception as e:
logger.error(f"Failed to list LoRAs: {str(e)}")
return []
def train_lora(
self,
dataset_name: str,
lora_name: str,
training_type: str = "vocal", # "vocal" or "symbolic"
config: Optional[Dict] = None,
progress_callback: Optional[Callable] = None
) -> Dict:
"""
Train LoRA adapter
Args:
dataset_name: Name of prepared dataset
lora_name: Name for the LoRA adapter
training_type: Type of training ("vocal" or "symbolic")
config: Training configuration (batch_size, learning_rate, etc.)
progress_callback: Optional callback for progress updates
Returns:
Training results dictionary
"""
try:
if self.is_training:
raise RuntimeError("Training already in progress")
self.is_training = True
logger.info(f"Starting LoRA training: {lora_name} ({training_type})")
# Load dataset
dataset_info = self.load_dataset(dataset_name)
if not dataset_info:
raise ValueError(f"Dataset not found: {dataset_name}")
# Check if dataset is from HuggingFace and needs preparation
if dataset_info.get('hf_dataset') and not dataset_info.get('prepared'):
raise ValueError(
f"Dataset '{dataset_name}' is a HuggingFace dataset that hasn't been prepared for training yet. "
f"Please use the 'User Audio Training' tab to upload and prepare your own audio files, "
f"or wait for dataset preparation features to be implemented."
)
# Validate dataset has required fields
if 'train_files' not in dataset_info or 'val_files' not in dataset_info:
raise ValueError(
f"Dataset '{dataset_name}' is missing required training files. "
f"Please use prepared datasets or upload your own audio in the 'User Audio Training' tab."
)
# Validate datasets are not empty
if not dataset_info['train_files'] or len(dataset_info['train_files']) == 0:
raise ValueError(
f"Dataset '{dataset_name}' has no training samples. "
f"The dataset may not have been prepared correctly. "
f"Please re-prepare the dataset or use a different one."
)
if not dataset_info['val_files'] or len(dataset_info['val_files']) == 0:
raise ValueError(
f"Dataset '{dataset_name}' has no validation samples. "
f"The dataset may not have been prepared correctly. "
f"Please re-prepare the dataset or use a different one."
)
# Default config
default_config = {
'batch_size': 4,
'learning_rate': 3e-4,
'num_epochs': 10,
'lora_rank': 16,
'lora_alpha': 32,
'warmup_steps': 100,
'save_every': 500,
'gradient_accumulation': 2
}
self.training_config = {**default_config, **(config or {})}
# Create datasets
train_dataset = TrainingDataset(
dataset_info['train_files'],
dataset_info['train_metadata']
)
val_dataset = TrainingDataset(
dataset_info['val_files'],
dataset_info['val_metadata']
)
# Create data loaders
# Disable pin_memory and num_workers for compatibility with ZeroGPU and CPU
# pin_memory requires persistent CUDA access which ZeroGPU doesn't provide at this stage
train_loader = DataLoader(
train_dataset,
batch_size=self.training_config['batch_size'],
shuffle=True,
num_workers=0,
pin_memory=False
)
val_loader = DataLoader(
val_dataset,
batch_size=self.training_config['batch_size'],
shuffle=False,
num_workers=0,
pin_memory=False
)
# Initialize model (placeholder - actual implementation would load DiffRhythm2)
# For now, we'll simulate training
logger.info("Initializing model and LoRA layers...")
# Note: Actual implementation would:
# 1. Load DiffRhythm2 model
# 2. Add LoRA adapters using peft library
# 3. Freeze base model, only train LoRA parameters
# Simulated training loop
num_steps = len(train_loader) * self.training_config['num_epochs']
logger.info(f"Training for {self.training_config['num_epochs']} epochs, {num_steps} total steps")
results = self._training_loop(
train_loader,
val_loader,
lora_name,
progress_callback
)
self.is_training = False
logger.info("Training complete!")
return results
except Exception as e:
self.is_training = False
logger.error(f"Training failed: {str(e)}")
raise
def _training_loop(
self,
train_loader: DataLoader,
val_loader: DataLoader,
lora_name: str,
progress_callback: Optional[Callable]
) -> Dict:
"""
Main training loop
Note: This is a simplified placeholder implementation.
Actual implementation would require:
1. Loading DiffRhythm2 model
2. Setting up LoRA adapters with peft library
3. Implementing proper loss functions
4. Gradient accumulation and optimization
"""
self.current_epoch = 0
self.current_step = 0
self.training_loss = []
best_val_loss = float('inf')
num_epochs = self.training_config['num_epochs']
for epoch in range(num_epochs):
self.current_epoch = epoch + 1
epoch_loss = 0.0
logger.info(f"Epoch {self.current_epoch}/{num_epochs}")
# Training phase
for batch_idx, batch in enumerate(train_loader):
self.current_step += 1
# Simulate training step
# Actual implementation would:
# 1. Move batch to device
# 2. Forward pass through model
# 3. Calculate loss
# 4. Backward pass
# 5. Update weights
# Simulated loss (decreasing over time)
step_loss = 1.0 / (1.0 + self.current_step * 0.01)
epoch_loss += step_loss
self.training_loss.append(step_loss)
# Progress update
if progress_callback and batch_idx % 10 == 0:
progress_callback({
'epoch': self.current_epoch,
'step': self.current_step,
'loss': step_loss,
'progress': (self.current_step / (len(train_loader) * num_epochs)) * 100
})
# Log every 50 steps
if self.current_step % 50 == 0:
logger.info(f"Step {self.current_step}: Loss = {step_loss:.4f}")
# Save checkpoint
if self.current_step % self.training_config['save_every'] == 0:
self._save_checkpoint(lora_name, self.current_step)
# Validation phase
avg_train_loss = epoch_loss / len(train_loader)
val_loss = self._validate(val_loader)
logger.info(f"Epoch {self.current_epoch}: Train Loss = {avg_train_loss:.4f}, Val Loss = {val_loss:.4f}")
# Save best model
if val_loss < best_val_loss:
best_val_loss = val_loss
self._save_lora_adapter(lora_name, is_best=True)
logger.info(f"New best model! Val Loss: {val_loss:.4f}")
# Final save
self._save_lora_adapter(lora_name, is_best=False)
return {
'lora_name': lora_name,
'num_epochs': num_epochs,
'total_steps': self.current_step,
'final_train_loss': avg_train_loss,
'final_val_loss': val_loss,
'best_val_loss': best_val_loss,
'training_time': 'simulated'
}
def _validate(self, val_loader: DataLoader) -> float:
"""Run validation"""
total_loss = 0.0
for batch in val_loader:
# Simulate validation
# Actual implementation would run model inference
val_loss = 1.0 / (1.0 + self.current_step * 0.01)
total_loss += val_loss
return total_loss / len(val_loader)
def _save_checkpoint(self, lora_name: str, step: int):
"""Save training checkpoint"""
checkpoint_dir = self.lora_dir / lora_name / "checkpoints"
checkpoint_dir.mkdir(parents=True, exist_ok=True)
checkpoint_path = checkpoint_dir / f"checkpoint_step_{step}.pt"
# Actual implementation would save:
# - LoRA weights
# - Optimizer state
# - Training step
# - Config
checkpoint_data = {
'step': step,
'epoch': self.current_epoch,
'config': self.training_config,
'loss_history': self.training_loss[-100:] # Last 100 steps
}
torch.save(checkpoint_data, checkpoint_path)
logger.info(f"Saved checkpoint: step_{step}")
def _save_lora_adapter(self, lora_name: str, is_best: bool = False):
"""Save final LoRA adapter"""
lora_path = self.lora_dir / lora_name
lora_path.mkdir(parents=True, exist_ok=True)
filename = "best_model.pt" if is_best else "final_model.pt"
save_path = lora_path / filename
# Actual implementation would save:
# - LoRA adapter weights only
# - Configuration
# - Training metadata
adapter_data = {
'lora_name': lora_name,
'config': self.training_config,
'training_steps': self.current_step,
'saved_at': datetime.now().isoformat()
}
torch.save(adapter_data, save_path)
logger.info(f"Saved LoRA adapter: {filename}")
# Save metadata
metadata_path = lora_path / "metadata.json"
with open(metadata_path, 'w') as f:
json.dump(adapter_data, f, indent=2)
def list_lora_adapters(self) -> List[Dict]:
"""List available LoRA adapters"""
try:
adapters = []
for lora_dir in self.lora_dir.iterdir():
if lora_dir.is_dir():
metadata_path = lora_dir / "metadata.json"
if metadata_path.exists():
with open(metadata_path, 'r') as f:
metadata = json.load(f)
adapters.append({
'name': lora_dir.name,
**metadata
})
else:
# Basic info if no metadata
adapters.append({
'name': lora_dir.name,
'has_best': (lora_dir / "best_model.pt").exists(),
'has_final': (lora_dir / "final_model.pt").exists()
})
return adapters
except Exception as e:
logger.error(f"Failed to list LoRA adapters: {str(e)}")
return []
def delete_lora_adapter(self, lora_name: str) -> bool:
"""Delete a LoRA adapter"""
try:
import shutil
lora_path = self.lora_dir / lora_name
if lora_path.exists():
shutil.rmtree(lora_path)
logger.info(f"Deleted LoRA adapter: {lora_name}")
return True
else:
logger.warning(f"LoRA adapter not found: {lora_name}")
return False
except Exception as e:
logger.error(f"Failed to delete LoRA adapter {lora_name}: {str(e)}")
return False
def stop_training(self):
"""Stop current training"""
if self.is_training:
logger.info("Training stop requested")
self.is_training = False
def get_training_status(self) -> Dict:
"""Get current training status"""
return {
'is_training': self.is_training,
'current_epoch': self.current_epoch,
'current_step': self.current_step,
'recent_loss': self.training_loss[-10:] if self.training_loss else [],
'config': self.training_config
}
def export_lora_adapter(self, lora_name: str) -> Optional[str]:
"""
Export a LoRA adapter as a zip file for download
Args:
lora_name: Name of the LoRA adapter to export
Returns:
Path to the exported zip file, or None if failed
"""
try:
import shutil
import tempfile
lora_path = self.lora_dir / lora_name
if not lora_path.exists():
logger.error(f"LoRA adapter not found: {lora_name}")
return None
# Create exports directory if it doesn't exist
exports_dir = Path("outputs/lora_exports")
exports_dir.mkdir(parents=True, exist_ok=True)
# Create zip file
zip_path = exports_dir / f"{lora_name}.zip"
# Remove existing zip if present
if zip_path.exists():
zip_path.unlink()
# Create zip archive
shutil.make_archive(
str(exports_dir / lora_name),
'zip',
str(lora_path)
)
logger.info(f"Exported LoRA adapter to: {zip_path}")
return str(zip_path)
except Exception as e:
logger.error(f"Failed to export LoRA adapter {lora_name}: {str(e)}")
return None
def import_lora_adapter(self, zip_path: str) -> Optional[str]:
"""
Import a LoRA adapter from a zip file
Args:
zip_path: Path to the zip file containing LoRA adapter
Returns:
Name of the imported LoRA adapter, or None if failed
"""
try:
import zipfile
import tempfile
# Extract to temporary directory first
with tempfile.TemporaryDirectory() as temp_dir:
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
zip_ref.extractall(temp_dir)
temp_path = Path(temp_dir)
# Case 1: Check if metadata.json is at root level
if (temp_path / "metadata.json").exists():
logger.info("Found metadata.json at root level")
source_dir = temp_path
# Read metadata to get LoRA name
with open(temp_path / "metadata.json", 'r') as f:
metadata = json.load(f)
lora_name = metadata.get('lora_name', 'imported_lora')
else:
# Case 2: Look for a subfolder with metadata.json
lora_folders = [d for d in temp_path.iterdir() if d.is_dir() and (d / "metadata.json").exists()]
if not lora_folders:
logger.error("No valid LoRA adapter found in zip file. Expected metadata.json at root or in a subfolder.")
return None
source_dir = lora_folders[0]
lora_name = source_dir.name
logger.info(f"Found LoRA in subfolder: {lora_name}")
# Copy to loras directory
dest_path = self.lora_dir / lora_name
# If already exists, rename with timestamp
if dest_path.exists():
timestamp = int(time.time())
original_name = lora_name
lora_name = f"{lora_name}_{timestamp}"
dest_path = self.lora_dir / lora_name
logger.info(f"LoRA '{original_name}' already exists, importing as '{lora_name}'")
# Create destination directory
dest_path.mkdir(parents=True, exist_ok=True)
# Copy all files from source to destination
for item in source_dir.iterdir():
if item.is_file():
shutil.copy2(item, dest_path / item.name)
elif item.is_dir():
shutil.copytree(item, dest_path / item.name)
logger.info(f"✅ Imported LoRA adapter: {lora_name}")
return lora_name
except Exception as e:
logger.error(f"Failed to import LoRA adapter: {str(e)}", exc_info=True)
return None
|