File size: 10,654 Bytes
156f8c8 | 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 | """
Training loop for the Thermal Pattern Analysis pipeline.
Supports:
- AdamW optimiser with cosine annealing scheduler
- Early stopping
- TensorBoard logging
- Checkpoint saving / resuming
- Mixed-precision training (if GPU available)
"""
import os
import time
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR
try:
from torch.utils.tensorboard import SummaryWriter
HAS_TENSORBOARD = True
except ImportError:
HAS_TENSORBOARD = False
SummaryWriter = None
from tqdm import tqdm
from pathlib import Path
from typing import Optional
from src.models.anomaly_detector import ThermalPatternPipeline
from src.training.losses import CombinedLoss
from src.evaluation.metrics import MetricsCalculator
class EarlyStopping:
"""Stop training when validation loss stops improving."""
def __init__(self, patience: int = 10, min_delta: float = 0.001):
self.patience = patience
self.min_delta = min_delta
self.counter = 0
self.best_loss = float("inf")
self.should_stop = False
def __call__(self, val_loss: float) -> bool:
if val_loss < self.best_loss - self.min_delta:
self.best_loss = val_loss
self.counter = 0
else:
self.counter += 1
if self.counter >= self.patience:
self.should_stop = True
return self.should_stop
class Trainer:
"""
Full training manager for the ThermalPatternPipeline.
"""
def __init__(
self,
model: ThermalPatternPipeline,
train_loader: DataLoader,
val_loader: DataLoader,
config,
device: torch.device,
):
self.model = model.to(device)
self.train_loader = train_loader
self.val_loader = val_loader
self.config = config
self.device = device
# Loss
self.criterion = CombinedLoss.from_config(config)
# Classification head (simple linear head for binary)
self.classifier = nn.Linear(
config.model.feature_extractor.embedding_dim, 2
).to(device)
# Optimiser: model params + classifier
all_params = list(model.parameters()) + list(self.classifier.parameters())
self.optimizer = AdamW(
all_params,
lr=config.training.learning_rate,
weight_decay=config.training.weight_decay,
)
# Scheduler
self.scheduler = CosineAnnealingLR(
self.optimizer,
T_max=config.training.epochs,
)
# Early stopping
es_cfg = config.training.early_stopping
self.early_stopping = EarlyStopping(
patience=es_cfg.patience,
min_delta=es_cfg.min_delta,
)
# Logging
log_dir = config.paths.get("logs", "logs")
if HAS_TENSORBOARD:
self.writer = SummaryWriter(log_dir=log_dir)
else:
self.writer = None
print(" ⚠ TensorBoard not available — logging to console only")
self.metrics = MetricsCalculator()
# Checkpoint dir
self.ckpt_dir = Path(config.paths.get("checkpoints", "checkpoints"))
self.ckpt_dir.mkdir(parents=True, exist_ok=True)
# Mixed-precision scaler
self.scaler = torch.amp.GradScaler("cuda") if device.type == "cuda" else None
def train_epoch(self, epoch: int) -> dict:
"""Run one training epoch."""
self.model.train()
self.classifier.train()
epoch_loss = 0.0
all_preds, all_labels = [], []
pbar = tqdm(self.train_loader, desc=f"Epoch {epoch+1} [Train]")
for sequences, labels in pbar:
sequences = sequences.to(self.device)
labels = labels.to(self.device)
self.optimizer.zero_grad()
# Forward
if self.scaler is not None:
with torch.amp.autocast("cuda"):
results = self.model(sequences)
logits = self.classifier(results["encoding"])
loss_dict = self.criterion(
results["encoding"], labels, logits
)
loss = loss_dict["total_loss"]
self.scaler.scale(loss).backward()
self.scaler.unscale_(self.optimizer)
nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
self.scaler.step(self.optimizer)
self.scaler.update()
else:
results = self.model(sequences)
logits = self.classifier(results["encoding"])
loss_dict = self.criterion(
results["encoding"], labels, logits
)
loss = loss_dict["total_loss"]
loss.backward()
nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
self.optimizer.step()
# Update baseline with normal samples
normal_mask = labels == 0
if normal_mask.any():
self.model.anomaly_detector.update_baseline(
results["encoding"][normal_mask].detach()
)
# Track metrics
epoch_loss += loss.item()
preds = logits.argmax(dim=1)
all_preds.extend(preds.cpu().numpy())
all_labels.extend(labels.cpu().numpy())
pbar.set_postfix(loss=f"{loss.item():.4f}")
avg_loss = epoch_loss / max(len(self.train_loader), 1)
metrics = self.metrics.compute_all(all_labels, all_preds)
metrics["loss"] = avg_loss
return metrics
@torch.no_grad()
def validate_epoch(self, epoch: int) -> dict:
"""Run one validation epoch."""
self.model.eval()
self.classifier.eval()
epoch_loss = 0.0
all_preds, all_labels, all_scores = [], [], []
for sequences, labels in tqdm(
self.val_loader, desc=f"Epoch {epoch+1} [Val]"
):
sequences = sequences.to(self.device)
labels = labels.to(self.device)
results = self.model(sequences)
logits = self.classifier(results["encoding"])
loss_dict = self.criterion(results["encoding"], labels, logits)
epoch_loss += loss_dict["total_loss"].item()
preds = logits.argmax(dim=1)
all_preds.extend(preds.cpu().numpy())
all_labels.extend(labels.cpu().numpy())
all_scores.extend(
torch.softmax(logits, dim=1)[:, 1].cpu().numpy()
)
avg_loss = epoch_loss / max(len(self.val_loader), 1)
metrics = self.metrics.compute_all(all_labels, all_preds, all_scores)
metrics["loss"] = avg_loss
return metrics
def train(self) -> dict:
"""
Full training loop with early stopping, checkpointing,
and TensorBoard logging.
Returns:
Best validation metrics dict.
"""
epochs = self.config.training.epochs
best_val_loss = float("inf")
best_metrics = {}
print(f"\n{'='*60}")
print(f" Training — {epochs} epochs on {self.device}")
print(f"{'='*60}\n")
for epoch in range(epochs):
t0 = time.time()
# Train
train_metrics = self.train_epoch(epoch)
# Validate
val_metrics = self.validate_epoch(epoch)
# Step scheduler
self.scheduler.step()
elapsed = time.time() - t0
# TensorBoard
if self.writer is not None:
for key, val in train_metrics.items():
self.writer.add_scalar(f"train/{key}", val, epoch)
for key, val in val_metrics.items():
self.writer.add_scalar(f"val/{key}", val, epoch)
self.writer.add_scalar(
"lr", self.optimizer.param_groups[0]["lr"], epoch
)
# Console summary
print(
f"Epoch {epoch+1:3d}/{epochs} | "
f"Train loss: {train_metrics['loss']:.4f} | "
f"Val loss: {val_metrics['loss']:.4f} | "
f"Val acc: {val_metrics.get('accuracy', 0):.4f} | "
f"Time: {elapsed:.1f}s"
)
# Checkpoint best model
if val_metrics["loss"] < best_val_loss:
best_val_loss = val_metrics["loss"]
best_metrics = val_metrics
self._save_checkpoint(epoch, val_metrics, is_best=True)
# Early stopping
if self.early_stopping(val_metrics["loss"]):
print(f"\n⏹ Early stopping at epoch {epoch+1}")
break
if self.writer is not None:
self.writer.close()
print(f"\n{'='*60}")
print(f" Training complete — Best val loss: {best_val_loss:.4f}")
print(f"{'='*60}\n")
return best_metrics
def _save_checkpoint(
self, epoch: int, metrics: dict, is_best: bool = False
):
"""Save model checkpoint."""
state = {
"epoch": epoch,
"model_state_dict": self.model.state_dict(),
"classifier_state_dict": self.classifier.state_dict(),
"optimizer_state_dict": self.optimizer.state_dict(),
"scheduler_state_dict": self.scheduler.state_dict(),
"metrics": metrics,
}
path = self.ckpt_dir / f"checkpoint_epoch_{epoch+1}.pt"
torch.save(state, path)
if is_best:
best_path = self.ckpt_dir / "best_model.pt"
torch.save(state, best_path)
def load_checkpoint(self, checkpoint_path: str):
"""Resume training from a saved checkpoint."""
ckpt = torch.load(checkpoint_path, map_location=self.device)
self.model.load_state_dict(ckpt["model_state_dict"])
self.classifier.load_state_dict(ckpt["classifier_state_dict"])
self.optimizer.load_state_dict(ckpt["optimizer_state_dict"])
self.scheduler.load_state_dict(ckpt["scheduler_state_dict"])
print(f"✓ Resumed from epoch {ckpt['epoch'] + 1}")
return ckpt["epoch"] + 1
|