Spaces:
Sleeping
Sleeping
File size: 16,760 Bytes
8f24287 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 | """
Real PyTorch Training Engine for the ML Training Optimizer environment.
Manages the full training lifecycle:
- Model creation with configurable architecture and dropout
- Optimizer creation (SGD, Adam, AdamW) with configurable hyperparameters
- Learning rate scheduling (constant, step, cosine, warmup_cosine)
- Data augmentation control
- Training loop execution
- Evaluation and metric tracking
- Best model checkpointing
"""
import math
import time
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Tuple
import torch
import torch.nn as nn
from torch.optim.lr_scheduler import CosineAnnealingLR, StepLR, LambdaLR
from .models_nn import create_model
from .datasets import create_dataloaders
@dataclass
class TrainingConfig:
"""Configuration for a training run, set by the agent."""
optimizer: str = "adam" # sgd, adam, adamw
learning_rate: float = 0.001
batch_size: int = 64
weight_decay: float = 0.0
dropout: float = 0.0
lr_schedule: str = "constant" # constant, step, cosine, warmup_cosine
warmup_epochs: int = 5
augmentation: bool = False
augmentation_strength: float = 0.5
@dataclass
class TrainingMetrics:
"""Metrics from a training run."""
train_loss: float = 0.0
val_loss: float = 0.0
train_accuracy: float = 0.0
val_accuracy: float = 0.0
current_lr: float = 0.0
epoch_time_seconds: float = 0.0
@dataclass
class TrainingState:
"""Full internal state of the trainer."""
current_epoch: int = 0
total_epochs_run: int = 0
best_val_accuracy: float = 0.0
best_val_epoch: int = 0
train_loss_history: List[float] = field(default_factory=list)
val_loss_history: List[float] = field(default_factory=list)
train_acc_history: List[float] = field(default_factory=list)
val_acc_history: List[float] = field(default_factory=list)
lr_history: List[float] = field(default_factory=list)
is_diverged: bool = False
config: Optional[TrainingConfig] = None
class Trainer:
"""
Manages real PyTorch model training on CPU.
Designed to be used by the environment: the agent configures hyperparameters,
then calls train_epochs() to actually train the model. The trainer tracks
all metrics and maintains the best model checkpoint.
"""
def __init__(
self,
model_type: str,
dataset_name: str,
max_epochs: int,
seed: int = 42,
):
self.model_type = model_type
self.dataset_name = dataset_name
self.max_epochs = max_epochs
self.seed = seed
# Set deterministic behavior
torch.manual_seed(seed)
torch.set_num_threads(2) # Match 2 vCPU constraint
if hasattr(torch, 'use_deterministic_algorithms'):
try:
torch.use_deterministic_algorithms(True)
except Exception:
pass # Some ops don't have deterministic impl
self.state = TrainingState()
self.model: Optional[nn.Module] = None
self.optimizer: Optional[torch.optim.Optimizer] = None
self.scheduler = None
self.train_loader = None
self.val_loader = None
self.best_model_state: Optional[dict] = None
self.device = torch.device("cpu")
self.criterion = nn.CrossEntropyLoss()
self._initialized = False
def configure(self, config: TrainingConfig) -> str:
"""
Configure (or reconfigure) the training setup.
Creates/rebuilds model, optimizer, scheduler, and data loaders
based on the provided configuration.
Returns:
Status message describing what was configured.
"""
old_config = self.state.config
self.state.config = config
# Rebuild model if dropout changed or first time
rebuild_model = (
not self._initialized
or old_config is None
or old_config.dropout != config.dropout
)
if rebuild_model:
torch.manual_seed(self.seed)
self.model = create_model(self.model_type, dropout=config.dropout)
self.model.to(self.device)
self.state.current_epoch = 0
self.state.total_epochs_run = 0
self.state.best_val_accuracy = 0.0
self.state.best_val_epoch = 0
self.state.train_loss_history = []
self.state.val_loss_history = []
self.state.train_acc_history = []
self.state.val_acc_history = []
self.state.lr_history = []
self.state.is_diverged = False
self.best_model_state = None
# Rebuild data loaders if batch size or augmentation changed
rebuild_data = (
not self._initialized
or old_config is None
or old_config.batch_size != config.batch_size
or old_config.augmentation != config.augmentation
or old_config.augmentation_strength != config.augmentation_strength
)
if rebuild_data:
self.train_loader, self.val_loader, _, _ = create_dataloaders(
self.dataset_name,
batch_size=config.batch_size,
seed=self.seed,
augment=config.augmentation,
aug_strength=config.augmentation_strength,
)
# Always rebuild optimizer (LR, weight decay, or optimizer type may change)
self._build_optimizer(config)
# Always rebuild scheduler
remaining = max(1, self.max_epochs - self.state.current_epoch)
self._build_scheduler(config, remaining)
self._initialized = True
status_parts = []
if rebuild_model:
param_count = sum(p.numel() for p in self.model.parameters())
status_parts.append(f"Model created: {self.model_type} ({param_count:,} params, dropout={config.dropout})")
if rebuild_data:
status_parts.append(f"Data loaded: {self.dataset_name} (batch_size={config.batch_size}, aug={'on' if config.augmentation else 'off'})")
status_parts.append(f"Optimizer: {config.optimizer} (lr={config.learning_rate}, wd={config.weight_decay})")
status_parts.append(f"Schedule: {config.lr_schedule}")
return " | ".join(status_parts)
def _build_optimizer(self, config: TrainingConfig):
"""Create optimizer from config."""
params = self.model.parameters()
if config.optimizer == "sgd":
self.optimizer = torch.optim.SGD(
params, lr=config.learning_rate,
weight_decay=config.weight_decay, momentum=0.9,
)
elif config.optimizer == "adam":
self.optimizer = torch.optim.Adam(
params, lr=config.learning_rate,
weight_decay=config.weight_decay,
)
elif config.optimizer == "adamw":
self.optimizer = torch.optim.AdamW(
params, lr=config.learning_rate,
weight_decay=config.weight_decay,
)
else:
raise ValueError(f"Unknown optimizer: {config.optimizer}")
def _build_scheduler(self, config: TrainingConfig, total_epochs: int):
"""Create learning rate scheduler from config."""
if config.lr_schedule == "constant":
self.scheduler = None
elif config.lr_schedule == "step":
self.scheduler = StepLR(self.optimizer, step_size=max(1, total_epochs // 3), gamma=0.1)
elif config.lr_schedule == "cosine":
self.scheduler = CosineAnnealingLR(self.optimizer, T_max=total_epochs)
elif config.lr_schedule == "warmup_cosine":
warmup = config.warmup_epochs
def lr_lambda(epoch):
if epoch < warmup:
return (epoch + 1) / warmup
progress = (epoch - warmup) / max(1, total_epochs - warmup)
return 0.5 * (1.0 + math.cos(math.pi * progress))
self.scheduler = LambdaLR(self.optimizer, lr_lambda)
else:
raise ValueError(f"Unknown LR schedule: {config.lr_schedule}")
def adjust_learning_rate(self, new_lr: float) -> str:
"""Adjust the learning rate on the live optimizer."""
if self.optimizer is None:
return "Error: No optimizer configured. Call configure_training first."
for param_group in self.optimizer.param_groups:
param_group['lr'] = new_lr
if self.state.config:
self.state.config.learning_rate = new_lr
return f"Learning rate adjusted to {new_lr}"
def train_epochs(self, num_epochs: int) -> List[TrainingMetrics]:
"""
Train for num_epochs epochs. Returns metrics for each epoch.
This runs REAL PyTorch training — forward pass, loss computation,
backward pass, optimizer step — on actual data.
"""
if not self._initialized or self.model is None:
raise RuntimeError("Trainer not configured. Call configure() first.")
if self.state.is_diverged:
return [TrainingMetrics(
train_loss=float('inf'), val_loss=float('inf'),
train_accuracy=0.0, val_accuracy=0.0,
current_lr=0.0, epoch_time_seconds=0.0,
)]
# Clamp to budget
remaining = self.max_epochs - self.state.current_epoch
actual_epochs = min(num_epochs, remaining)
if actual_epochs <= 0:
return []
all_metrics = []
for _ in range(actual_epochs):
epoch_start = time.time()
# --- Training phase ---
self.model.train()
running_loss = 0.0
correct = 0
total = 0
for inputs, targets in self.train_loader:
inputs, targets = inputs.to(self.device), targets.to(self.device)
self.optimizer.zero_grad()
outputs = self.model(inputs)
loss = self.criterion(outputs, targets)
loss.backward()
self.optimizer.step()
running_loss += loss.item() * inputs.size(0)
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
train_loss = running_loss / total
train_acc = correct / total
# Check for divergence
if math.isnan(train_loss) or math.isinf(train_loss) or train_loss > 100:
self.state.is_diverged = True
metrics = TrainingMetrics(
train_loss=float('inf'), val_loss=float('inf'),
train_accuracy=0.0, val_accuracy=0.0,
current_lr=self.optimizer.param_groups[0]['lr'],
epoch_time_seconds=time.time() - epoch_start,
)
all_metrics.append(metrics)
break
# --- Validation phase ---
val_loss, val_acc = self._evaluate()
# Step scheduler
if self.scheduler is not None:
self.scheduler.step()
current_lr = self.optimizer.param_groups[0]['lr']
# Update state
self.state.current_epoch += 1
self.state.total_epochs_run += 1
self.state.train_loss_history.append(train_loss)
self.state.val_loss_history.append(val_loss)
self.state.train_acc_history.append(train_acc)
self.state.val_acc_history.append(val_acc)
self.state.lr_history.append(current_lr)
# Track best model
if val_acc > self.state.best_val_accuracy:
self.state.best_val_accuracy = val_acc
self.state.best_val_epoch = self.state.current_epoch
self.best_model_state = {
k: v.clone() for k, v in self.model.state_dict().items()
}
epoch_time = time.time() - epoch_start
metrics = TrainingMetrics(
train_loss=train_loss,
val_loss=val_loss,
train_accuracy=train_acc,
val_accuracy=val_acc,
current_lr=current_lr,
epoch_time_seconds=epoch_time,
)
all_metrics.append(metrics)
return all_metrics
def _evaluate(self) -> Tuple[float, float]:
"""Evaluate model on validation set. Returns (val_loss, val_accuracy)."""
self.model.eval()
running_loss = 0.0
correct = 0
total = 0
with torch.no_grad():
for inputs, targets in self.val_loader:
inputs, targets = inputs.to(self.device), targets.to(self.device)
outputs = self.model(inputs)
loss = self.criterion(outputs, targets)
running_loss += loss.item() * inputs.size(0)
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
return running_loss / total, correct / total
def get_metrics_summary(self) -> Dict:
"""Get current training metrics as a dictionary."""
if not self.state.train_loss_history:
return {
"current_epoch": 0,
"max_epochs": self.max_epochs,
"remaining_budget": self.max_epochs,
"train_loss": 0.0,
"val_loss": 0.0,
"train_accuracy": 0.0,
"val_accuracy": 0.0,
"best_val_accuracy": 0.0,
"best_val_epoch": 0,
"loss_history_last_10": [],
"val_loss_history_last_10": [],
"convergence_signal": "not_started",
"is_diverged": False,
}
# Determine convergence signal
signal = self._get_convergence_signal()
return {
"current_epoch": self.state.current_epoch,
"max_epochs": self.max_epochs,
"remaining_budget": self.max_epochs - self.state.current_epoch,
"train_loss": round(self.state.train_loss_history[-1], 6),
"val_loss": round(self.state.val_loss_history[-1], 6),
"train_accuracy": round(self.state.train_acc_history[-1], 4),
"val_accuracy": round(self.state.val_acc_history[-1], 4),
"best_val_accuracy": round(self.state.best_val_accuracy, 4),
"best_val_epoch": self.state.best_val_epoch,
"loss_history_last_10": [round(x, 4) for x in self.state.train_loss_history[-10:]],
"val_loss_history_last_10": [round(x, 4) for x in self.state.val_loss_history[-10:]],
"convergence_signal": signal,
"is_diverged": self.state.is_diverged,
}
def _get_convergence_signal(self) -> str:
"""Analyze recent training dynamics to produce a human-readable signal."""
if self.state.is_diverged:
return "diverged"
if len(self.state.val_loss_history) < 3:
return "warming_up"
recent_val = self.state.val_loss_history[-5:]
recent_train = self.state.train_loss_history[-5:]
# Check overfitting: train improving but val getting worse
if len(recent_val) >= 3:
val_trend = recent_val[-1] - recent_val[0]
train_trend = recent_train[-1] - recent_train[0]
if val_trend > 0 and train_trend < 0:
return "overfitting"
# Check plateau: val loss barely changing
if len(recent_val) >= 3:
val_range = max(recent_val) - min(recent_val)
if val_range < 0.005:
return "plateaued"
# Check if improving
if recent_val[-1] < recent_val[0]:
return "improving"
return "stalling"
def get_overfitting_gap(self) -> float:
"""Returns train_accuracy - val_accuracy (overfitting indicator)."""
if not self.state.train_acc_history:
return 0.0
return self.state.train_acc_history[-1] - self.state.val_acc_history[-1]
def get_wasted_epochs(self) -> int:
"""Epochs trained after best val accuracy was achieved."""
if self.state.best_val_epoch == 0:
return 0
return max(0, self.state.current_epoch - self.state.best_val_epoch)
def get_loss_variance(self, window: int = 10) -> float:
"""Variance of recent validation losses."""
if len(self.state.val_loss_history) < 2:
return 0.0
recent = self.state.val_loss_history[-window:]
mean = sum(recent) / len(recent)
return sum((x - mean) ** 2 for x in recent) / len(recent)
|