Spaces:
Build error
Build error
File size: 12,558 Bytes
583e46a | 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 | """
Unified Trainer class that orchestrates the full training pipeline.
Takes model, dataloaders, config → trains, evaluates, saves artifacts.
"""
import sys
import time
from pathlib import Path
from typing import Optional
import torch
import torch.nn as nn
PROJECT_ROOT = Path(__file__).resolve().parents[3]
sys.path.insert(0, str(PROJECT_ROOT))
from ml.src.training.engine import train_one_epoch, validate, measure_inference_latency
from ml.src.training.losses import create_loss
from ml.src.training.optimizer_factory import create_optimizer
from ml.src.training.scheduler_factory import create_scheduler
from ml.src.training.callbacks import EarlyStopping, ModelCheckpoint, MetricLogger
from ml.src.evaluation.metrics import compute_metrics
from ml.src.evaluation.plots import (
plot_training_curves,
plot_confusion_matrix,
plot_per_class_f1,
)
from ml.src.utils.io import save_json, ensure_dirs
from ml.src.utils.seed import set_seed
class Trainer:
"""
Unified training orchestrator.
Usage:
trainer = Trainer(model, config, dataloaders, class_names, device)
trainer.train()
trainer.evaluate()
trainer.save_artifacts()
"""
def __init__(
self,
model: nn.Module,
config: dict,
dataloaders: dict,
class_names: list[str],
device: torch.device,
model_name: Optional[str] = None,
):
self.model = model.to(device)
self.config = config
self.dataloaders = dataloaders
self.class_names = class_names
self.device = device
self.model_name = model_name or config.get('model', {}).get('name', 'model')
# Training config
train_cfg = config.get('training', {})
self.num_epochs = train_cfg.get('num_epochs', 30)
self.gradient_clip = train_cfg.get('gradient_clip_max_norm', 1.0)
# Setup artifact directories
artifacts_cfg = config.get('artifacts', {})
self.checkpoint_dir = PROJECT_ROOT / artifacts_cfg.get('checkpoints_dir', 'ml/artifacts/checkpoints')
self.figures_dir = PROJECT_ROOT / artifacts_cfg.get('figures_dir', 'ml/artifacts/figures')
self.logs_dir = PROJECT_ROOT / artifacts_cfg.get('logs_dir', 'ml/artifacts/logs')
self.reports_dir = PROJECT_ROOT / artifacts_cfg.get('reports_dir', 'ml/artifacts/reports')
ensure_dirs(self.checkpoint_dir, self.figures_dir, self.logs_dir, self.reports_dir)
# Create loss
label_smoothing = train_cfg.get('label_smoothing', 0.0)
self.criterion = create_loss(label_smoothing=label_smoothing)
# Create optimizer
self.optimizer = create_optimizer(
self.model,
optimizer_name=train_cfg.get('optimizer', 'adam'),
learning_rate=train_cfg.get('learning_rate', 0.001),
weight_decay=train_cfg.get('weight_decay', 0.0001),
)
# Create scheduler
self.scheduler = create_scheduler(
self.optimizer,
scheduler_name=train_cfg.get('scheduler', 'cosine'),
num_epochs=self.num_epochs,
step_size=train_cfg.get('scheduler_step_size', 10),
gamma=train_cfg.get('scheduler_gamma', 0.1),
warmup_epochs=train_cfg.get('warmup_epochs', 0),
)
# Callbacks
self.early_stopping = EarlyStopping(
patience=train_cfg.get('early_stopping_patience', 7),
mode='min',
)
self.checkpoint = ModelCheckpoint(
save_dir=self.checkpoint_dir,
model_name=self.model_name,
mode='min',
)
self.logger = MetricLogger(
log_dir=self.logs_dir,
model_name=self.model_name,
)
# Results storage
self.test_results = None
self.training_summary = None
def train(self) -> dict:
"""Run the full training loop."""
print(f"\n{'='*60}")
print(f"Training: {self.model_name}")
print(f"Device: {self.device}")
print(f"Epochs: {self.num_epochs}")
print(f"{'='*60}\n")
set_seed(self.config.get('data', {}).get('seed', 42))
for epoch in range(1, self.num_epochs + 1):
# Train
train_metrics = train_one_epoch(
self.model, self.dataloaders['train'], self.criterion,
self.optimizer, self.device, self.gradient_clip,
)
# Validate
val_metrics = validate(
self.model, self.dataloaders['val'], self.criterion, self.device,
)
# Get current LR
current_lr = self.optimizer.param_groups[0]['lr']
# Log
self.logger.log_epoch(epoch, train_metrics, val_metrics, current_lr)
# Checkpoint
self.checkpoint(
val_metrics['loss'], self.model, self.optimizer, epoch,
extra_info={'config': self.config, 'class_names': self.class_names},
)
# Step scheduler
if self.scheduler is not None:
if isinstance(self.scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
self.scheduler.step(val_metrics['loss'])
else:
self.scheduler.step()
# Early stopping
if self.early_stopping(val_metrics['loss']):
print(f"\nEarly stopping at epoch {epoch}")
break
# Save training history
self.logger.save()
self.training_summary = self.logger.get_summary()
print(f"\nTraining complete! Best val loss: {self.checkpoint.best_value:.4f}")
return self.training_summary
def evaluate(self, split: str = 'test') -> dict:
"""
Evaluate the best model on test set.
Loads the best checkpoint and runs full evaluation.
"""
print(f"\n{'='*60}")
print(f"Evaluating: {self.model_name} on {split}")
print(f"{'='*60}\n")
# Load best checkpoint
if self.checkpoint.best_path and self.checkpoint.best_path.exists():
ckpt = torch.load(self.checkpoint.best_path, map_location=self.device, weights_only=False)
self.model.load_state_dict(ckpt['model_state_dict'])
print(f"Loaded best checkpoint from epoch {ckpt.get('epoch', '?')}")
# Run evaluation
results = validate(
self.model, self.dataloaders[split], self.criterion, self.device,
)
# Compute detailed metrics
metrics = compute_metrics(
results['labels'], results['predictions'], self.class_names,
)
# Measure inference latency
latency = measure_inference_latency(
self.model, self.device,
img_size=self.config.get('image', {}).get('size', 224),
)
# Calculate model size
model_size_mb = sum(
p.numel() * p.element_size() for p in self.model.parameters()
) / (1024 * 1024)
self.test_results = {
'model_name': self.model_name,
'split': split,
'loss': results['loss'],
'metrics': metrics,
'latency': latency,
'model_size_mb': round(model_size_mb, 2),
'num_parameters': sum(p.numel() for p in self.model.parameters()),
}
# Print summary
print(f"\n{split.upper()} Results:")
print(f" Loss: {results['loss']:.4f}")
print(f" Accuracy: {metrics['accuracy']:.4f}")
print(f" Macro F1: {metrics['macro_f1']:.4f}")
print(f" Macro Prec: {metrics['macro_precision']:.4f}")
print(f" Macro Recall: {metrics['macro_recall']:.4f}")
print(f" Avg Latency: {latency['avg_ms']:.2f} ms")
print(f" Model Size: {model_size_mb:.2f} MB")
return self.test_results
def save_artifacts(self) -> None:
"""Save all evaluation artifacts: plots, reports."""
print(f"\nSaving artifacts for {self.model_name}...")
model_fig_dir = self.figures_dir / self.model_name
model_fig_dir.mkdir(parents=True, exist_ok=True)
# Training curves
plot_training_curves(
self.logger.history,
save_path=model_fig_dir / 'training_curves.png',
title=f'{self.model_name} - Training Curves',
)
# Confusion matrix (from test results)
if self.test_results:
labels = self.test_results.get('_raw_labels', None)
preds = self.test_results.get('_raw_predictions', None)
# Load from last validation run if needed
if labels is None:
val_results = validate(
self.model, self.dataloaders.get('test', self.dataloaders['val']),
self.criterion, self.device,
)
labels = val_results['labels']
preds = val_results['predictions']
plot_confusion_matrix(
labels, preds, self.class_names,
save_path=model_fig_dir / 'confusion_matrix.png',
title=f'{self.model_name} - Confusion Matrix',
)
# Per-class F1
if self.test_results and 'metrics' in self.test_results:
per_class = self.test_results['metrics'].get('per_class', {})
if per_class:
plot_per_class_f1(
per_class, self.class_names,
save_path=model_fig_dir / 'per_class_f1.png',
title=f'{self.model_name} - Per-Class F1 Score',
)
# Save detailed report JSON
if self.test_results:
report = {**self.test_results}
report.pop('_raw_labels', None)
report.pop('_raw_predictions', None)
save_json(report, self.reports_dir / f'{self.model_name}_report.json')
# Save training summary
if self.training_summary:
save_json(
self.training_summary,
self.reports_dir / f'{self.model_name}_training_summary.json',
)
print(f" Artifacts saved to {self.figures_dir / self.model_name}")
def train_with_phase_switch(
self,
phase1_epochs: int = 10,
phase2_lr: float = 0.0001,
unfreeze_layers: list[str] = None,
) -> dict:
"""
Two-phase training for transfer learning models.
Phase 1: Frozen backbone, train head only.
Phase 2: Unfreeze specified layers, use lower LR.
"""
print(f"\n{'='*60}")
print(f"Phase 1: Frozen backbone training ({phase1_epochs} epochs)")
print(f"{'='*60}\n")
# Phase 1
original_epochs = self.num_epochs
self.num_epochs = phase1_epochs
self.train()
# Phase 2: Unfreeze
remaining_epochs = original_epochs - phase1_epochs
if remaining_epochs > 0 and hasattr(self.model, 'unfreeze_layers'):
print(f"\n{'='*60}")
print(f"Phase 2: Fine-tuning ({remaining_epochs} epochs)")
print(f"{'='*60}\n")
self.model.unfreeze_layers(unfreeze_layers)
# Use param groups if available
if hasattr(self.model, 'get_param_groups'):
base_lr = self.config.get('training', {}).get('learning_rate', 0.001)
param_groups = self.model.get_param_groups(base_lr, phase2_lr)
self.optimizer = create_optimizer(
param_groups,
optimizer_name=self.config.get('training', {}).get('optimizer', 'adamw'),
learning_rate=base_lr,
weight_decay=self.config.get('training', {}).get('weight_decay', 0.0001),
)
else:
for pg in self.optimizer.param_groups:
pg['lr'] = phase2_lr
# Reset scheduler and early stopping
self.scheduler = create_scheduler(
self.optimizer,
scheduler_name=self.config.get('training', {}).get('scheduler', 'cosine'),
num_epochs=remaining_epochs,
)
self.early_stopping = EarlyStopping(
patience=self.config.get('training', {}).get('early_stopping_patience', 7),
)
self.num_epochs = remaining_epochs
self.train()
self.num_epochs = original_epochs
return self.logger.get_summary()
|