ACE-Step-Custom / src /lora_trainer.py
ACE-Step Custom
Deploy ACE-Step Custom Edition with bug fixes
a602628
"""
LoRA Trainer - Handles LoRA training for custom models
"""
import torch
import torchaudio
from pathlib import Path
import logging
from typing import List, Dict, Any, Optional, Callable
import json
from datetime import datetime
logger = logging.getLogger(__name__)
class LoRATrainer:
"""Manages LoRA training for ACE-Step model."""
def __init__(self, config: Dict[str, Any]):
"""
Initialize LoRA trainer.
Args:
config: Configuration dictionary
"""
self.config = config
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.training_dir = Path(config.get("training_dir", "lora_training"))
self.training_dir.mkdir(exist_ok=True)
self.model = None
self.lora_config = None
logger.info(f"LoRA Trainer initialized on {self.device}")
def prepare_dataset(self, audio_files: List[str]) -> List[str]:
"""
Prepare audio files for training.
Args:
audio_files: List of audio file paths
Returns:
List of prepared file paths
"""
try:
logger.info(f"Preparing {len(audio_files)} files for training...")
prepared_dir = self.training_dir / "prepared_data" / datetime.now().strftime("%Y%m%d_%H%M%S")
prepared_dir.mkdir(parents=True, exist_ok=True)
prepared_files = []
for i, file_path in enumerate(audio_files):
try:
# Load audio
audio, sr = torchaudio.load(file_path)
# Resample to target sample rate if needed
target_sr = self.config.get("sample_rate", 44100)
if sr != target_sr:
resampler = torchaudio.transforms.Resample(sr, target_sr)
audio = resampler(audio)
# Convert to mono if needed (for some training scenarios)
if audio.shape[0] > 1 and self.config.get("force_mono", False):
audio = torch.mean(audio, dim=0, keepdim=True)
# Normalize
audio = audio / (torch.abs(audio).max() + 1e-8)
# Split long files into chunks if needed
chunk_duration = self.config.get("chunk_duration", 30) # seconds
chunk_samples = int(chunk_duration * target_sr)
if audio.shape[1] > chunk_samples:
# Split into chunks
num_chunks = audio.shape[1] // chunk_samples
for j in range(num_chunks):
start = j * chunk_samples
end = start + chunk_samples
chunk = audio[:, start:end]
# Save chunk
chunk_path = prepared_dir / f"audio_{i:04d}_chunk_{j:02d}.wav"
torchaudio.save(
str(chunk_path),
chunk,
target_sr,
encoding="PCM_S",
bits_per_sample=16
)
prepared_files.append(str(chunk_path))
else:
# Save as-is
output_path = prepared_dir / f"audio_{i:04d}.wav"
torchaudio.save(
str(output_path),
audio,
target_sr,
encoding="PCM_S",
bits_per_sample=16
)
prepared_files.append(str(output_path))
except Exception as e:
logger.warning(f"Failed to process {file_path}: {e}")
continue
# Save dataset metadata
metadata = {
"num_files": len(prepared_files),
"original_files": len(audio_files),
"sample_rate": target_sr,
"prepared_at": datetime.now().isoformat(),
"files": prepared_files
}
metadata_path = prepared_dir / "metadata.json"
with open(metadata_path, 'w') as f:
json.dump(metadata, f, indent=2)
logger.info(f"✅ Prepared {len(prepared_files)} training files")
return prepared_files
except Exception as e:
logger.error(f"Dataset preparation failed: {e}")
raise
def initialize_lora(self, rank: int = 16, alpha: int = 32):
"""
Initialize LoRA configuration.
Args:
rank: LoRA rank
alpha: LoRA alpha
"""
try:
from peft import LoraConfig, get_peft_model
self.lora_config = LoraConfig(
r=rank,
lora_alpha=alpha,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"], # Attention layers
lora_dropout=0.1,
bias="none",
task_type="CAUSAL_LM"
)
logger.info(f"✅ LoRA initialized: rank={rank}, alpha={alpha}")
except Exception as e:
logger.error(f"LoRA initialization failed: {e}")
raise
def load_lora(self, lora_path: str):
"""
Load existing LoRA model for continued training.
Args:
lora_path: Path to LoRA model
"""
try:
from peft import PeftModel
from transformers import AutoModel
# Load base model
base_model = AutoModel.from_pretrained(
self.config.get("model_path", "ACE-Step/ACE-Step-v1-3.5B"),
torch_dtype=torch.float16 if self.device.type == "cuda" else torch.float32
)
# Load with LoRA
self.model = PeftModel.from_pretrained(base_model, lora_path)
logger.info(f"✅ Loaded LoRA from {lora_path}")
except Exception as e:
logger.error(f"Failed to load LoRA: {e}")
raise
def train(
self,
dataset_path: str,
model_name: str,
learning_rate: float = 1e-4,
batch_size: int = 4,
num_epochs: int = 10,
progress_callback: Optional[Callable] = None
) -> str:
"""
Train LoRA model.
Args:
dataset_path: Path to prepared dataset
model_name: Name for the trained model
learning_rate: Learning rate
batch_size: Batch size
num_epochs: Number of epochs
progress_callback: Optional callback for progress updates
Returns:
Path to trained model
"""
try:
logger.info(f"Starting LoRA training: {model_name}")
# Load dataset
dataset = self._load_dataset(dataset_path)
# Load base model if not already loaded
if self.model is None:
from transformers import AutoModel
from peft import get_peft_model
base_model = AutoModel.from_pretrained(
self.config.get("model_path", "ACE-Step/ACE-Step-v1-3.5B"),
torch_dtype=torch.float16 if self.device.type == "cuda" else torch.float32,
device_map="auto"
)
self.model = get_peft_model(base_model, self.lora_config)
self.model.train()
# Setup optimizer
optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=learning_rate,
weight_decay=0.01
)
# Training loop
total_steps = (len(dataset) // batch_size) * num_epochs
step = 0
for epoch in range(num_epochs):
epoch_loss = 0.0
for batch_idx in range(0, len(dataset), batch_size):
batch = dataset[batch_idx:batch_idx + batch_size]
# Forward pass (simplified - actual implementation would be more complex)
loss = self._training_step(batch)
# Backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
epoch_loss += loss.item()
step += 1
# Progress callback
if progress_callback:
progress_callback(step, total_steps, loss.item())
avg_loss = epoch_loss / (len(dataset) // batch_size)
logger.info(f"Epoch {epoch+1}/{num_epochs} - Loss: {avg_loss:.4f}")
# Save trained model
output_dir = self.training_dir / "models" / model_name
output_dir.mkdir(parents=True, exist_ok=True)
self.model.save_pretrained(str(output_dir))
# Save training info
info = {
"model_name": model_name,
"learning_rate": learning_rate,
"batch_size": batch_size,
"num_epochs": num_epochs,
"dataset_size": len(dataset),
"trained_at": datetime.now().isoformat(),
"lora_config": {
"rank": self.lora_config.r,
"alpha": self.lora_config.lora_alpha
}
}
info_path = output_dir / "training_info.json"
with open(info_path, 'w') as f:
json.dump(info, f, indent=2)
logger.info(f"✅ Training complete! Model saved to {output_dir}")
return str(output_dir)
except Exception as e:
logger.error(f"Training failed: {e}")
raise
def _load_dataset(self, dataset_path: str) -> List[Dict[str, Any]]:
"""Load prepared dataset."""
dataset_path = Path(dataset_path)
# Load metadata
metadata_path = dataset_path / "metadata.json"
if metadata_path.exists():
with open(metadata_path, 'r') as f:
metadata = json.load(f)
files = metadata.get("files", [])
else:
# Scan directory for audio files
files = list(dataset_path.glob("*.wav"))
dataset = []
for file_path in files:
dataset.append({
"path": str(file_path),
"audio": None # Lazy loading
})
return dataset
def _training_step(self, batch: List[Dict[str, Any]]) -> torch.Tensor:
"""
Perform single training step.
This is a simplified placeholder - actual implementation would:
1. Load audio from batch
2. Encode to latent space
3. Generate predictions
4. Calculate loss
5. Return loss
Args:
batch: Training batch
Returns:
Loss tensor
"""
# Placeholder loss calculation
# Actual implementation would process audio through model
loss = torch.tensor(0.5, requires_grad=True, device=self.device)
return loss
def export_for_inference(self, lora_path: str, output_path: str):
"""
Export LoRA model for inference.
Args:
lora_path: Path to LoRA model
output_path: Output path for exported model
"""
try:
# Load LoRA
self.load_lora(lora_path)
# Merge LoRA with base model
merged_model = self.model.merge_and_unload()
# Save merged model
merged_model.save_pretrained(output_path)
logger.info(f"✅ Exported model to {output_path}")
except Exception as e:
logger.error(f"Export failed: {e}")
raise