voice-model-rl-training / voice_rl /models /voice_model_wrapper.py
mbellan's picture
Initial deployment
c3efd49
"""Voice model wrapper for HuggingFace models."""
import torch
import torch.nn as nn
import logging
from typing import Optional, Iterator, Dict, Any, Tuple
from pathlib import Path
from transformers import AutoModel, AutoConfig, AutoProcessor
import json
from .policy_wrapper import RLVoiceModel
logger = logging.getLogger(__name__)
class VoiceModelWrapper:
"""
Wrapper for HuggingFace voice models with RL training support.
Provides a consistent interface for model loading, inference,
checkpointing, and license verification.
"""
# List of known commercial-use licenses
COMMERCIAL_LICENSES = [
"apache-2.0",
"mit",
"bsd",
"bsd-3-clause",
"cc-by-4.0",
"cc-by-sa-4.0",
"openrail",
]
def __init__(
self,
model_name: str,
device: str = "cuda",
cache_dir: Optional[str] = None,
enable_rl: bool = True,
action_dim: int = 256
):
"""
Initialize the voice model wrapper.
Args:
model_name: HuggingFace model identifier
device: Device to load model on ('cuda', 'cpu', 'mps')
cache_dir: Optional cache directory for model files
enable_rl: Whether to add RL policy/value heads
action_dim: Dimensionality of action space for RL
"""
self.model_name = model_name
self.device = device
self.cache_dir = cache_dir
self.enable_rl = enable_rl
self.action_dim = action_dim
self.model = None
self.rl_model = None
self.processor = None
self.config = None
logger.info(f"Initialized VoiceModelWrapper for {model_name} on {device} (RL: {enable_rl})")
def load_model(self) -> None:
"""
Load the voice model from HuggingFace.
Performs license verification and architecture compatibility checks.
Raises:
ValueError: If model has incompatible license or architecture
RuntimeError: If model loading fails
"""
try:
logger.info(f"Loading model: {self.model_name}")
# Load configuration first
self.config = AutoConfig.from_pretrained(
self.model_name,
cache_dir=self.cache_dir
)
# Verify license
self._verify_license()
# Verify architecture compatibility
self._verify_architecture()
# Load model
self.model = AutoModel.from_pretrained(
self.model_name,
cache_dir=self.cache_dir
)
self.model.to(self.device)
self.model.train() # Set to training mode for RL
# Wrap with RL policy/value heads if enabled
if self.enable_rl:
hidden_size = self.config.hidden_size if hasattr(self.config, 'hidden_size') else 768
self.rl_model = RLVoiceModel(
base_model=self.model,
hidden_size=hidden_size,
action_dim=self.action_dim
)
self.rl_model.to(self.device)
logger.info(f"Added RL policy/value heads (action_dim={self.action_dim})")
# Load processor if available
try:
self.processor = AutoProcessor.from_pretrained(
self.model_name,
cache_dir=self.cache_dir
)
except Exception as e:
logger.warning(f"Could not load processor: {e}")
self.processor = None
logger.info(f"Successfully loaded model: {self.model_name}")
logger.info(f"Model parameters: {self.count_parameters():,}")
except Exception as e:
error_msg = f"Failed to load model {self.model_name}: {str(e)}"
logger.error(error_msg)
raise RuntimeError(error_msg) from e
def _verify_license(self) -> None:
"""
Verify that the model has a commercial-use license.
Raises:
ValueError: If license is not suitable for commercial use
"""
# Try to get license from config
license_info = getattr(self.config, 'license', None)
if license_info is None:
logger.warning(
f"No license information found for {self.model_name}. "
"Please verify license manually."
)
return
license_lower = license_info.lower()
# Check if license is in approved list
is_commercial = any(
approved in license_lower
for approved in self.COMMERCIAL_LICENSES
)
if not is_commercial:
raise ValueError(
f"Model {self.model_name} has license '{license_info}' "
f"which may not be suitable for commercial use. "
f"Approved licenses: {', '.join(self.COMMERCIAL_LICENSES)}"
)
logger.info(f"License verified: {license_info}")
def _verify_architecture(self) -> None:
"""
Verify that the model architecture is compatible with RL training.
Checks for required attributes and methods.
Raises:
ValueError: If architecture is incompatible
"""
# Check if model has required architecture attributes
required_attrs = ['config']
for attr in required_attrs:
if not hasattr(self.config, attr.replace('config.', '')):
logger.warning(f"Model may be missing attribute: {attr}")
# Check model type
model_type = getattr(self.config, 'model_type', 'unknown')
logger.info(f"Model type: {model_type}")
# Verify model can be put in training mode
if self.model is not None and not hasattr(self.model, 'train'):
raise ValueError("Model does not support training mode")
logger.info("Architecture compatibility verified")
def generate(
self,
input_features: torch.Tensor,
training: bool = False,
**kwargs
) -> torch.Tensor:
"""
Generate output from the model.
Args:
input_features: Input tensor
training: If True, compute with gradients (for RL training)
**kwargs: Additional generation parameters
Returns:
Generated output tensor
Raises:
RuntimeError: If model is not loaded
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
if training:
# During training, keep gradients for backprop
outputs = self.model(input_features, **kwargs)
else:
# During inference, no gradients needed
with torch.no_grad():
outputs = self.model(input_features, **kwargs)
# Handle different output types
if hasattr(outputs, 'last_hidden_state'):
return outputs.last_hidden_state
elif isinstance(outputs, torch.Tensor):
return outputs
else:
return outputs[0]
def get_logits(self, input_features: torch.Tensor) -> torch.Tensor:
"""
Get model logits for input features.
Args:
input_features: Input tensor
Returns:
Logits tensor
Raises:
RuntimeError: If model is not loaded
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
outputs = self.model(input_features)
if hasattr(outputs, 'logits'):
return outputs.logits
elif hasattr(outputs, 'last_hidden_state'):
return outputs.last_hidden_state
else:
return outputs[0]
def forward(self, input_features: torch.Tensor, **kwargs) -> Any:
"""
Forward pass through the model.
Args:
input_features: Input tensor
**kwargs: Additional forward parameters
Returns:
Model outputs (RL-compatible if RL enabled)
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
# Use RL model if available (returns log_probs, values)
if self.rl_model is not None:
return self.rl_model(input_features, **kwargs)
else:
return self.model(input_features, **kwargs)
def sample_action(
self,
input_features: torch.Tensor,
deterministic: bool = False
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Sample action from the policy (RL training).
Args:
input_features: Input audio features
deterministic: If True, take most likely action
Returns:
Tuple of (actions, log_probs, values)
Raises:
RuntimeError: If RL model is not enabled
"""
if self.rl_model is None:
raise RuntimeError("RL model not enabled. Set enable_rl=True when initializing.")
return self.rl_model.sample_action(input_features, deterministic)
def evaluate_actions(
self,
input_features: torch.Tensor,
actions: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Evaluate actions (for PPO training).
Args:
input_features: Input audio features
actions: Actions to evaluate
Returns:
Tuple of (log_probs, values, entropy)
Raises:
RuntimeError: If RL model is not enabled
"""
if self.rl_model is None:
raise RuntimeError("RL model not enabled. Set enable_rl=True when initializing.")
return self.rl_model.evaluate_actions(input_features, actions)
def save_checkpoint(self, path: str, metadata: Optional[Dict] = None) -> None:
"""
Save model checkpoint.
Args:
path: Path to save checkpoint
metadata: Optional metadata to save with checkpoint
Raises:
RuntimeError: If model is not loaded
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
checkpoint_path = Path(path)
checkpoint_path.parent.mkdir(parents=True, exist_ok=True)
checkpoint = {
'model_state_dict': self.model.state_dict(),
'model_name': self.model_name,
'config': self.config.to_dict() if self.config else None,
'enable_rl': self.enable_rl,
'action_dim': self.action_dim,
}
# Save RL model state if present
if self.rl_model is not None:
checkpoint['rl_model_state_dict'] = self.rl_model.state_dict()
if metadata:
checkpoint['metadata'] = metadata
torch.save(checkpoint, checkpoint_path)
logger.info(f"Checkpoint saved to {checkpoint_path}")
def load_checkpoint(self, path: str) -> Dict:
"""
Load model checkpoint.
Args:
path: Path to checkpoint file
Returns:
Checkpoint metadata
Raises:
RuntimeError: If model is not loaded
FileNotFoundError: If checkpoint file doesn't exist
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
checkpoint_path = Path(path)
if not checkpoint_path.exists():
raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")
checkpoint = torch.load(checkpoint_path, map_location=self.device)
self.model.load_state_dict(checkpoint['model_state_dict'])
# Load RL model state if present
if 'rl_model_state_dict' in checkpoint and self.rl_model is not None:
self.rl_model.load_state_dict(checkpoint['rl_model_state_dict'])
logger.info("Loaded RL model state")
logger.info(f"Checkpoint loaded from {checkpoint_path}")
return checkpoint.get('metadata', {})
def get_trainable_parameters(self) -> Iterator[torch.nn.Parameter]:
"""
Get iterator over trainable parameters.
Returns:
Iterator over trainable parameters
Raises:
RuntimeError: If model is not loaded
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
return (p for p in self.model.parameters() if p.requires_grad)
def count_parameters(self, trainable_only: bool = False) -> int:
"""
Count model parameters.
Args:
trainable_only: If True, count only trainable parameters
Returns:
Number of parameters
"""
if self.model is None:
return 0
# Count RL model params if available, otherwise base model
model_to_count = self.rl_model if self.rl_model is not None else self.model
if trainable_only:
return sum(p.numel() for p in model_to_count.parameters() if p.requires_grad)
else:
return sum(p.numel() for p in model_to_count.parameters())
def set_training_mode(self, mode: bool = True) -> None:
"""
Set model training mode.
Args:
mode: If True, set to training mode; otherwise evaluation mode
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
if mode:
self.model.train()
if self.rl_model is not None:
self.rl_model.train()
else:
self.model.eval()
if self.rl_model is not None:
self.rl_model.eval()
def to(self, device: str) -> None:
"""
Move model to specified device.
Args:
device: Target device
"""
if self.model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
self.device = device
self.model.to(device)
if self.rl_model is not None:
self.rl_model.to(device)
logger.info(f"Model moved to {device}")
def get_rl_model(self) -> Optional[nn.Module]:
"""
Get the RL-wrapped model.
Returns:
RLVoiceModel if RL is enabled, None otherwise
"""
return self.rl_model