File size: 9,660 Bytes
8d18b7c | 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 | """Checkpoint Management for Training"""
import json
import logging
import shutil
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any, Dict, List, Optional
import torch
logger = logging.getLogger(__name__)
@dataclass
class CheckpointMetadata:
"""Metadata for a checkpoint."""
step: int
epoch: int
global_step: int
metrics: Dict[str, float] = field(default_factory=dict)
config: Dict[str, Any] = field(default_factory=dict)
model_name: str = "zenith"
timestamp: str = ""
def to_dict(self) -> Dict[str, Any]:
return {
"step": self.step,
"epoch": self.epoch,
"global_step": self.global_step,
"metrics": self.metrics,
"config": self.config,
"model_name": self.model_name,
"timestamp": self.timestamp,
}
@classmethod
def from_dict(cls, data: Dict[str, Any]) -> "CheckpointMetadata":
return cls(**data)
class CheckpointManager:
"""Manages saving and loading of checkpoints."""
def __init__(
self,
checkpoint_dir: str,
save_total_limit: int = 5,
save_best_only: bool = False,
metric_for_best: str = "eval_loss",
greater_is_better: bool = False,
):
self.checkpoint_dir = Path(checkpoint_dir)
self.checkpoint_dir.mkdir(parents=True, exist_ok=True)
self.save_total_limit = save_total_limit
self.save_best_only = save_best_only
self.metric_for_best = metric_for_best
self.greater_is_better = greater_is_better
self.best_metric = None
self.checkpoints: List[Path] = []
# Load existing checkpoints
self._scan_checkpoints()
def _scan_checkpoints(self):
"""Scan checkpoint directory for existing checkpoints."""
for path in self.checkpoint_dir.glob("checkpoint-*"):
if path.is_dir():
self.checkpoints.append(path)
self.checkpoints.sort(key=lambda p: int(p.name.split("-")[1]))
def save_checkpoint(
self,
state: Dict[str, Any],
name: str,
metrics: Optional[Dict[str, float]] = None,
) -> Path:
"""Save checkpoint to disk."""
checkpoint_path = self.checkpoint_dir / f"checkpoint-{name}"
checkpoint_path.mkdir(exist_ok=True)
# Save model state
torch.save(state["model_state_dict"], checkpoint_path / "pytorch_model.bin")
# Save optimizer and scheduler states
if "optimizer_state_dict" in state:
torch.save(state["optimizer_state_dict"], checkpoint_path / "optimizer.pt")
if "scheduler_state_dict" in state and state["scheduler_state_dict"]:
torch.save(state["scheduler_state_dict"], checkpoint_path / "scheduler.pt")
if "scaler_state_dict" in state and state["scaler_state_dict"]:
torch.save(state["scaler_state_dict"], checkpoint_path / "scaler.pt")
# Save metadata
metadata = CheckpointMetadata(
step=state.get("step", 0),
epoch=state.get("epoch", 0),
global_step=state.get("global_step", 0),
metrics=metrics or {},
config=state.get("config", {}),
timestamp=state.get("timestamp", ""),
)
with open(checkpoint_path / "metadata.json", "w") as f:
json.dump(metadata.to_dict(), f, indent=2)
logger.info(f"Checkpoint saved: {checkpoint_path}")
# Update checkpoint list
if checkpoint_path not in self.checkpoints:
self.checkpoints.append(checkpoint_path)
self.checkpoints.sort(key=lambda p: int(p.name.split("-")[1]))
# Enforce limit
if self.save_total_limit > 0 and len(self.checkpoints) > self.save_total_limit:
self._remove_oldest_checkpoint()
return checkpoint_path
def load_checkpoint(
self,
checkpoint_path: Union[str, Path],
model: torch.nn.Module,
optimizer: Optional[torch.optim.Optimizer] = None,
scheduler: Optional[Any] = None,
scaler: Optional[torch.cuda.amp.GradScaler] = None,
) -> CheckpointMetadata:
"""Load checkpoint from disk."""
checkpoint_path = Path(checkpoint_path)
if not checkpoint_path.exists():
raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")
# Load model
model_path = checkpoint_path / "pytorch_model.bin"
if model_path.exists():
state_dict = torch.load(model_path, map_location="cpu")
model.load_state_dict(state_dict)
logger.info(f"Loaded model from {model_path}")
else:
logger.warning(f"Model weights not found at {model_path}")
# Load optimizer
if optimizer is not None:
opt_path = checkpoint_path / "optimizer.pt"
if opt_path.exists():
optimizer.load_state_dict(torch.load(opt_path, map_location="cpu"))
logger.info(f"Loaded optimizer from {opt_path}")
# Load scheduler
if scheduler is not None:
sched_path = checkpoint_path / "scheduler.pt"
if sched_path.exists():
scheduler.load_state_dict(torch.load(sched_path, map_location="cpu"))
logger.info(f"Loaded scheduler from {sched_path}")
# Load scaler
if scaler is not None:
scaler_path = checkpoint_path / "scaler.pt"
if scaler_path.exists():
scaler.load_state_dict(torch.load(scaler_path, map_location="cpu"))
logger.info(f"Loaded scaler from {scaler_path}")
# Load metadata
meta_path = checkpoint_path / "metadata.json"
if meta_path.exists():
with open(meta_path, "r") as f:
metadata = CheckpointMetadata.from_dict(json.load(f))
logger.info(f"Loaded metadata: epoch={metadata.epoch}, step={metadata.step}")
else:
metadata = CheckpointMetadata(step=0, epoch=0, global_step=0)
return metadata
def get_latest_checkpoint(self) -> Optional[Path]:
"""Get the most recent checkpoint."""
if self.checkpoints:
return self.checkpoints[-1]
return None
def get_best_checkpoint(self) -> Optional[Path]:
"""Get the best checkpoint based on metric."""
if not self.checkpoints:
return None
best_path = None
best_value = None
for path in self.checkpoints:
meta_path = path / "metadata.json"
if meta_path.exists():
with open(meta_path, "r") as f:
meta = CheckpointMetadata.from_dict(json.load(f))
if self.metric_for_best in meta.metrics:
value = meta.metrics[self.metric_for_best]
if best_value is None or (
self.greater_is_better and value > best_value
) or (not self.greater_is_better and value < best_value):
best_value = value
best_path = path
return best_path
def _remove_oldest_checkpoint(self):
"""Remove the oldest checkpoint to maintain limit."""
if len(self.checkpoints) > self.save_total_limit:
oldest = self.checkpoints.pop(0)
if oldest.exists():
shutil.rmtree(oldest)
logger.info(f"Removed old checkpoint: {oldest}")
def cleanup(self, keep: Optional[List[Path]] = None):
"""Clean up checkpoints, optionally keeping specific ones."""
if keep is None:
keep = []
for path in self.checkpoints:
if path not in keep:
if path.exists():
shutil.rmtree(path)
logger.info(f"Removed checkpoint: {path}")
self._scan_checkpoints()
def save_checkpoint(
model: torch.nn.Module,
optimizer: torch.optim.Optimizer,
scheduler: Optional[Any],
scaler: Optional[torch.cuda.amp.GradScaler],
checkpoint_dir: str,
epoch: int,
global_step: int,
metrics: Optional[Dict[str, float]] = None,
config: Optional[Dict[str, Any]] = None,
save_optimizer: bool = True,
save_scheduler: bool = True,
):
"""Convenience function to save a checkpoint."""
manager = CheckpointManager(checkpoint_dir, save_total_limit=0)
state = {
"model_state_dict": model.state_dict(),
"global_step": global_step,
"epoch": epoch,
"config": config or {},
"timestamp": "",
}
if save_optimizer:
state["optimizer_state_dict"] = optimizer.state_dict()
if save_scheduler and scheduler is not None:
state["scheduler_state_dict"] = scheduler.state_dict()
manager.save_checkpoint(state, f"step-{global_step}", metrics)
def load_checkpoint(
checkpoint_path: str,
model: torch.nn.Module,
optimizer: Optional[torch.optim.Optimizer] = None,
scheduler: Optional[Any] = None,
scaler: Optional[torch.cuda.amp.GradScaler] = None,
) -> int:
"""Convenience function to load a checkpoint."""
manager = CheckpointManager(Path(checkpoint_path).parent)
metadata = manager.load_checkpoint(checkpoint_path, model, optimizer, scheduler, scaler)
return metadata.global_step
|