File size: 12,311 Bytes
c8b77b5 | 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 | """
Loss Functions
=============
This module implements various loss functions for neural network training,
including cross-entropy, KL divergence, and custom loss functions for
the MangoMAS multi-agent system.
"""
import logging
from abc import ABC, abstractmethod
from typing import Dict, Optional, Any
import torch
import torch.nn.functional as F
logger = logging.getLogger(__name__)
class LossFunction(ABC):
"""Abstract base class for all loss functions"""
def __init__(self, reduction: str = 'mean'):
self.reduction = reduction
@abstractmethod
def forward(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""Compute the loss"""
pass
def __call__(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""Call the loss function"""
return self.forward(predictions, targets)
class CrossEntropyLoss(LossFunction):
"""
Cross-entropy loss for classification tasks
Mathematical formulation:
L = -∑(y_i * log(ŷ_i))
Where y_i is the true label and ŷ_i is the predicted probability.
"""
def __init__(self, reduction: str = 'mean', label_smoothing: float = 0.0,
weight: Optional[torch.Tensor] = None):
super().__init__(reduction)
self.label_smoothing = label_smoothing
self.weight = weight
def forward(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Compute cross-entropy loss
Args:
predictions: Model predictions (logits)
targets: True labels
Returns:
Computed loss
"""
# Apply softmax to get probabilities
if predictions.dim() > 1 and predictions.size(1) > 1:
# Multi-class classification
log_probs = F.log_softmax(predictions, dim=1)
if targets.dim() == 1:
# Targets are class indices
loss = F.nll_loss(log_probs, targets, weight=self.weight,
reduction=self.reduction, label_smoothing=self.label_smoothing)
else:
# Targets are one-hot encoded
loss = -(targets * log_probs).sum(dim=1)
if self.reduction == 'mean':
loss = loss.mean()
elif self.reduction == 'sum':
loss = loss.sum()
else:
# Binary classification
loss = F.binary_cross_entropy_with_logits(predictions, targets.float(),
weight=self.weight, reduction=self.reduction)
return loss
class KLDivergenceLoss(LossFunction):
"""
Kullback-Leibler divergence loss for distribution matching
Mathematical formulation:
KL(P||Q) = ∑ P(x) * log(P(x)/Q(x))
"""
def __init__(self, reduction: str = 'mean', log_target: bool = False):
super().__init__(reduction)
self.log_target = log_target
def forward(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Compute KL divergence loss
Args:
predictions: Predicted distribution (log probabilities)
targets: Target distribution (probabilities or log probabilities)
Returns:
Computed KL divergence loss
"""
if self.log_target:
# Both predictions and targets are in log space
loss = F.kl_div(predictions, targets, reduction=self.reduction, log_target=True)
else:
# Convert predictions to log space, targets are probabilities
log_predictions = F.log_softmax(predictions, dim=-1)
loss = F.kl_div(log_predictions, targets, reduction=self.reduction, log_target=False)
return loss
class MSELoss(LossFunction):
"""
Mean Squared Error loss for regression tasks
Mathematical formulation:
L = (1/n) * ∑(y_i - ŷ_i)²
"""
def forward(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Compute MSE loss
Args:
predictions: Model predictions
targets: True values
Returns:
Computed MSE loss
"""
loss = F.mse_loss(predictions, targets, reduction=self.reduction)
return loss
class HuberLoss(LossFunction):
"""
Huber loss (smooth L1 loss) for robust regression
Mathematical formulation:
L = { 0.5 * (y - ŷ)², if |y - ŷ| < δ
{ δ * (|y - ŷ| - 0.5 * δ), otherwise
"""
def __init__(self, reduction: str = 'mean', delta: float = 1.0):
super().__init__(reduction)
self.delta = delta
def forward(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Compute Huber loss
Args:
predictions: Model predictions
targets: True values
Returns:
Computed Huber loss
"""
loss = F.huber_loss(predictions, targets, reduction=self.reduction, delta=self.delta)
return loss
class FocalLoss(LossFunction):
"""
Focal loss for addressing class imbalance
Mathematical formulation:
FL = -α(1-p_t)^γ * log(p_t)
Where p_t is the predicted probability for the true class.
"""
def __init__(self, alpha: float = 1.0, gamma: float = 2.0, reduction: str = 'mean'):
super().__init__(reduction)
self.alpha = alpha
self.gamma = gamma
def forward(self, predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Compute focal loss
Args:
predictions: Model predictions (logits)
targets: True labels
Returns:
Computed focal loss
"""
# Compute probabilities
probs = F.softmax(predictions, dim=1)
# Get probabilities for true classes
if targets.dim() == 1:
# Targets are class indices
pt = probs.gather(1, targets.unsqueeze(1)).squeeze(1)
else:
# Targets are one-hot encoded
pt = (probs * targets).sum(dim=1)
# Compute focal loss
focal_weight = self.alpha * (1 - pt) ** self.gamma
loss = -focal_weight * torch.log(pt + 1e-8)
if self.reduction == 'mean':
loss = loss.mean()
elif self.reduction == 'sum':
loss = loss.sum()
return loss
class ContrastiveLoss(LossFunction):
"""
Contrastive loss for learning representations
Mathematical formulation:
L = (1-y) * d² + y * max(0, margin - d)²
Where d is the distance between embeddings and y is the similarity label.
"""
def __init__(self, margin: float = 1.0, reduction: str = 'mean'):
super().__init__(reduction)
self.margin = margin
def forward(self, embeddings1: torch.Tensor, embeddings2: torch.Tensor,
labels: torch.Tensor) -> torch.Tensor:
"""
Compute contrastive loss
Args:
embeddings1: First set of embeddings
embeddings2: Second set of embeddings
labels: Similarity labels (1 for similar, 0 for dissimilar)
Returns:
Computed contrastive loss
"""
# Compute Euclidean distance
distance = F.pairwise_distance(embeddings1, embeddings2)
# Compute contrastive loss
positive_loss = labels.float() * distance.pow(2)
negative_loss = (1 - labels.float()) * F.relu(self.margin - distance).pow(2)
loss = positive_loss + negative_loss
if self.reduction == 'mean':
loss = loss.mean()
elif self.reduction == 'sum':
loss = loss.sum()
return loss
class CustomAgentLoss(LossFunction):
"""
Custom loss function for MangoMAS agents
Combines multiple loss components to optimize agent performance.
"""
def __init__(self, task_loss_weight: float = 1.0,
consistency_loss_weight: float = 0.1,
regularization_weight: float = 0.01):
super().__init__('mean')
self.task_loss_weight = task_loss_weight
self.consistency_loss_weight = consistency_loss_weight
self.regularization_weight = regularization_weight
# Initialize component losses
self.task_loss = CrossEntropyLoss()
self.consistency_loss = MSELoss()
def forward(self, predictions: torch.Tensor, targets: torch.Tensor,
model_outputs: Optional[Dict[str, torch.Tensor]] = None) -> torch.Tensor:
"""
Compute custom agent loss
Args:
predictions: Model predictions
targets: True labels
model_outputs: Additional model outputs for consistency loss
Returns:
Computed custom loss
"""
# Task-specific loss
task_loss = self.task_loss(predictions, targets)
# Consistency loss (if model outputs provided)
consistency_loss = torch.tensor(0.0, device=predictions.device)
if model_outputs is not None and 'hidden_states' in model_outputs:
# Encourage consistent hidden representations
hidden_states = model_outputs['hidden_states']
if len(hidden_states) > 1:
# Compute consistency between consecutive hidden states
for i in range(len(hidden_states) - 1):
consistency_loss += self.consistency_loss(hidden_states[i], hidden_states[i+1])
consistency_loss /= (len(hidden_states) - 1)
# Regularization loss (L2 penalty)
regularization_loss = torch.tensor(0.0, device=predictions.device)
# This would be computed from model parameters in practice
# Combine losses
total_loss = (self.task_loss_weight * task_loss +
self.consistency_loss_weight * consistency_loss +
self.regularization_weight * regularization_loss)
return total_loss
class LossFunctionFactory:
"""Factory class for creating loss functions"""
@staticmethod
def create_loss_function(loss_type: str, **kwargs) -> LossFunction:
"""Create a loss function instance"""
loss_functions = {
'cross_entropy': CrossEntropyLoss,
'kl_divergence': KLDivergenceLoss,
'mse': MSELoss,
'huber': HuberLoss,
'focal': FocalLoss,
'contrastive': ContrastiveLoss,
'custom_agent': CustomAgentLoss
}
if loss_type.lower() not in loss_functions:
raise ValueError(f"Unknown loss function type: {loss_type}")
loss_class = loss_functions[loss_type.lower()]
return loss_class(**kwargs)
@staticmethod
def get_default_config(loss_type: str) -> Dict[str, Any]:
"""Get default configuration for loss function"""
configs = {
'cross_entropy': {
'reduction': 'mean',
'label_smoothing': 0.0
},
'kl_divergence': {
'reduction': 'mean',
'log_target': False
},
'mse': {
'reduction': 'mean'
},
'huber': {
'reduction': 'mean',
'delta': 1.0
},
'focal': {
'alpha': 1.0,
'gamma': 2.0,
'reduction': 'mean'
},
'contrastive': {
'margin': 1.0,
'reduction': 'mean'
},
'custom_agent': {
'task_loss_weight': 1.0,
'consistency_loss_weight': 0.1,
'regularization_weight': 0.01
}
}
return configs.get(loss_type.lower(), {})
|