Spaces:
Runtime error
Runtime error
File size: 8,199 Bytes
ec8f374 |
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 |
"""
Custom Training Loop Module
Provides custom training loop implementation with fine-grained control over training.
"""
from dataclasses import dataclass, field
from typing import Optional, Dict, Any, Callable
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
@dataclass
class TrainingConfig:
"""Configuration for custom training loop."""
num_epochs: int = 3
learning_rate: float = 2e-4
batch_size: int = 4
gradient_accumulation_steps: int = 4
max_grad_norm: float = 1.0
warmup_steps: int = 100
logging_steps: int = 10
eval_steps: int = 500
save_steps: int = 500
output_dir: str = "./models/output"
device: str = "cuda" if torch.cuda.is_available() else "cpu"
class TrainingLoop:
"""
Custom training loop for fine-grained control over the training process.
Provides manual control over:
- Forward/backward passes
- Gradient accumulation
- Learning rate scheduling
- Logging and evaluation
- Checkpointing
"""
def __init__(
self,
model: torch.nn.Module,
train_dataloader: DataLoader,
eval_dataloader: Optional[DataLoader] = None,
config: Optional[TrainingConfig] = None
):
"""
Initialize custom training loop.
Args:
model: PyTorch model to train
train_dataloader: Training data loader
eval_dataloader: Optional evaluation data loader
config: Training configuration
"""
self.model = model
self.train_dataloader = train_dataloader
self.eval_dataloader = eval_dataloader
self.config = config or TrainingConfig()
self.optimizer = None
self.scheduler = None
self.global_step = 0
self.current_epoch = 0
def setup_optimizer(self, optimizer_class=torch.optim.AdamW, **optimizer_kwargs):
"""
Setup optimizer and learning rate scheduler.
Args:
optimizer_class: Optimizer class to use
**optimizer_kwargs: Additional optimizer arguments
"""
self.optimizer = optimizer_class(
self.model.parameters(),
lr=self.config.learning_rate,
**optimizer_kwargs
)
# Linear warmup scheduler
def lr_lambda(current_step: int):
if current_step < self.config.warmup_steps:
return float(current_step) / float(max(1, self.config.warmup_steps))
return 1.0
self.scheduler = torch.optim.lr_scheduler.LambdaLR(
self.optimizer,
lr_lambda
)
def train_step(self, batch: Dict[str, torch.Tensor]) -> float:
"""
Perform a single training step.
Args:
batch: Batch of training data
Returns:
Loss value
"""
# Move batch to device
batch = {k: v.to(self.config.device) for k, v in batch.items()}
# Forward pass
outputs = self.model(**batch)
loss = outputs.loss
# Scale loss for gradient accumulation
loss = loss / self.config.gradient_accumulation_steps
# Backward pass
loss.backward()
return loss.item()
def train_epoch(self) -> Dict[str, float]:
"""
Train for one epoch.
Returns:
Training metrics
"""
self.model.train()
total_loss = 0
num_batches = 0
progress_bar = tqdm(
self.train_dataloader,
desc=f"Epoch {self.current_epoch + 1}/{self.config.num_epochs}"
)
for step, batch in enumerate(progress_bar):
# Training step
loss = self.train_step(batch)
total_loss += loss
# Gradient accumulation
if (step + 1) % self.config.gradient_accumulation_steps == 0:
# Clip gradients
torch.nn.utils.clip_grad_norm_(
self.model.parameters(),
self.config.max_grad_norm
)
# Optimizer step
self.optimizer.step()
self.scheduler.step()
self.optimizer.zero_grad()
self.global_step += 1
num_batches += 1
# Update progress bar
progress_bar.set_postfix({
"loss": total_loss / num_batches,
"lr": self.scheduler.get_last_lr()[0]
})
# Logging
if self.global_step % self.config.logging_steps == 0:
avg_loss = total_loss / num_batches
print(f"Step {self.global_step}: loss={avg_loss:.4f}")
# Evaluation
if self.eval_dataloader and self.global_step % self.config.eval_steps == 0:
eval_metrics = self.evaluate()
print(f"Evaluation: {eval_metrics}")
self.model.train()
return {
"loss": total_loss / max(num_batches, 1),
"epoch": self.current_epoch
}
def evaluate(self) -> Dict[str, float]:
"""
Evaluate model on validation set.
Returns:
Evaluation metrics
"""
if self.eval_dataloader is None:
return {}
self.model.eval()
total_loss = 0
num_batches = 0
with torch.no_grad():
for batch in tqdm(self.eval_dataloader, desc="Evaluating"):
batch = {k: v.to(self.config.device) for k, v in batch.items()}
outputs = self.model(**batch)
total_loss += outputs.loss.item()
num_batches += 1
return {
"eval_loss": total_loss / max(num_batches, 1)
}
def train(self, callback: Optional[Callable] = None) -> Dict[str, Any]:
"""
Run full training loop.
Args:
callback: Optional callback function called after each epoch
Returns:
Training history
"""
if self.optimizer is None:
self.setup_optimizer()
print(f"Starting training for {self.config.num_epochs} epochs")
print(f"Device: {self.config.device}")
print(f"Batch size: {self.config.batch_size}")
print(f"Gradient accumulation: {self.config.gradient_accumulation_steps}")
history = {
"train_loss": [],
"eval_loss": []
}
for epoch in range(self.config.num_epochs):
self.current_epoch = epoch
# Train epoch
train_metrics = self.train_epoch()
history["train_loss"].append(train_metrics["loss"])
# Evaluate
if self.eval_dataloader:
eval_metrics = self.evaluate()
history["eval_loss"].append(eval_metrics.get("eval_loss", 0))
# Callback
if callback:
callback(epoch, train_metrics)
print("✅ Training complete!")
return history
def save_checkpoint(self, path: str) -> None:
"""
Save training checkpoint.
Args:
path: Path to save checkpoint
"""
checkpoint = {
"model_state_dict": self.model.state_dict(),
"optimizer_state_dict": self.optimizer.state_dict(),
"scheduler_state_dict": self.scheduler.state_dict(),
"global_step": self.global_step,
"epoch": self.current_epoch
}
torch.save(checkpoint, path)
print(f"Checkpoint saved to: {path}")
def load_checkpoint(self, path: str) -> None:
"""
Load training checkpoint.
Args:
path: Path to checkpoint
"""
checkpoint = torch.load(path)
self.model.load_state_dict(checkpoint["model_state_dict"])
self.optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
self.scheduler.load_state_dict(checkpoint["scheduler_state_dict"])
self.global_step = checkpoint["global_step"]
self.current_epoch = checkpoint["epoch"]
print(f"Checkpoint loaded from: {path}")
|