forked from MichalDanielDobrzanski/DeepLearningPython
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtraining.py
More file actions
383 lines (342 loc) · 14.2 KB
/
Copy pathtraining.py
File metadata and controls
383 lines (342 loc) · 14.2 KB
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
"""Training loop and classroom demo orchestration."""
from __future__ import annotations
import json
import math
import time
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, Iterable, Optional
import torch
from rich.console import Console
from rich.live import Live
from torch import nn
from torch.optim import Optimizer
from .checkpoints import load_checkpoint, save_checkpoint
from .data import DataConfig, get_mnist_dataloaders, iter_batch_preview
from .metrics import MetricsTracker
from .models import ModelConfig
from .visualization import LivePreview, TrainingDashboard
@dataclass
class TrainingConfig:
epochs: int = 5
optimizer: str = "sgd"
learning_rate: float = 0.1
momentum: float = 0.9
weight_decay: float = 0.0
batch_size: int = 64
val_batch_size: Optional[int] = None
device: str = "auto"
checkpoint_dir: Path = Path("artifacts/checkpoints")
checkpoint_interval: int = 200
log_dir: Path = Path("artifacts/logs")
metrics_filename: str = "metrics.jsonl"
resume_from: Optional[Path] = None
preview_interval: int = 25
log_interval: int = 10
evaluate_every: int = 1
limit_train_batches: Optional[int] = None
limit_val_batches: Optional[int] = None
limit_train_samples: Optional[int] = None
limit_val_samples: Optional[int] = None
mixed_precision: bool = False
gradient_clip: Optional[float] = None
scheduler: str = "none"
scheduler_step_size: int = 1
scheduler_gamma: float = 0.95
num_workers: int = 0
seed: int = 0
enable_live: bool = True
def resolve_device(self) -> torch.device:
if self.device == "auto":
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
return torch.device(self.device)
@property
def log_path(self) -> Path:
return Path(self.log_dir) / self.metrics_filename
def build_optimizer(model: nn.Module, config: TrainingConfig) -> Optimizer:
params = [p for p in model.parameters() if p.requires_grad]
if config.optimizer == "sgd":
return torch.optim.SGD(
params,
lr=config.learning_rate,
momentum=config.momentum,
weight_decay=config.weight_decay,
)
if config.optimizer == "adam":
return torch.optim.Adam(
params,
lr=config.learning_rate,
weight_decay=config.weight_decay,
)
raise ValueError(f"Unsupported optimizer: {config.optimizer}")
def build_scheduler(
optimizer: Optimizer, config: TrainingConfig
) -> Optional[torch.optim.lr_scheduler._LRScheduler]:
if config.scheduler == "none":
return None
if config.scheduler == "steplr":
return torch.optim.lr_scheduler.StepLR(
optimizer,
step_size=config.scheduler_step_size,
gamma=config.scheduler_gamma,
)
if config.scheduler == "cosine":
return torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
T_max=max(config.epochs, 1),
)
raise ValueError(f"Unsupported scheduler: {config.scheduler}")
def compute_accuracy(predictions: torch.Tensor, labels: torch.Tensor) -> float:
predicted = predictions.argmax(dim=1)
return float((predicted == labels).float().mean().item())
def compute_gradient_norm(model: nn.Module) -> float:
total = 0.0
for param in model.parameters():
if param.grad is None:
continue
total += param.grad.detach().data.norm(2).item() ** 2
return math.sqrt(total)
class Trainer:
"""High-level orchestration for the interactive classroom trainer."""
def __init__(
self,
*,
model_config: ModelConfig,
data_config: DataConfig,
training_config: TrainingConfig,
console: Optional[Console] = None,
) -> None:
self.console = console or Console()
self.model_config = model_config
self.data_config = data_config
self.training_config = training_config
self.device = self.training_config.resolve_device()
self.dashboard = TrainingDashboard(console=self.console)
self.metrics_tracker = MetricsTracker()
self.global_step = 0
self.log_path = self.training_config.log_path
self.log_path.parent.mkdir(parents=True, exist_ok=True)
self._log_file = self.log_path.open("w", encoding="utf8")
# region lifecycle helpers
def _load_data(self):
data_config = DataConfig(
data_dir=self.data_config.data_dir,
batch_size=self.training_config.batch_size,
val_batch_size=self.training_config.val_batch_size,
download=self.data_config.download,
num_workers=self.training_config.num_workers,
limit_train_samples=self.training_config.limit_train_samples,
limit_val_samples=self.training_config.limit_val_samples,
seed=self.training_config.seed,
use_fake_data=self.data_config.use_fake_data,
)
return get_mnist_dataloaders(data_config)
def _create_model(self) -> nn.Module:
model = self.model_config.create_model()
model.to(self.device)
return model
def _write_metrics_log(self, payload: Dict[str, object]) -> None:
json.dump(payload, self._log_file)
self._log_file.write("\n")
self._log_file.flush()
def _restore_checkpoint(
self, model: nn.Module, optimizer: Optimizer, scheduler
) -> Dict[str, int]:
if not self.training_config.resume_from:
return {"step": 0, "epoch": 0}
state = load_checkpoint(self.training_config.resume_from, map_location=self.device)
model.load_state_dict(state.model_state)
optimizer.load_state_dict(state.optimizer_state)
if scheduler and state.scheduler_state:
scheduler.load_state_dict(state.scheduler_state)
self.console.print(f"Resumed from {self.training_config.resume_from} at step {state.step}")
return {"step": state.step, "epoch": state.epoch}
def _maybe_checkpoint(
self,
*,
model: nn.Module,
optimizer: Optimizer,
scheduler,
epoch: int,
step: int,
) -> None:
interval = self.training_config.checkpoint_interval
if interval <= 0 or step % interval != 0:
return
checkpoint_dir = Path(self.training_config.checkpoint_dir)
checkpoint_dir.mkdir(parents=True, exist_ok=True)
path = checkpoint_dir / f"step_{step:07d}.pt"
save_checkpoint(
path=path,
step=step,
epoch=epoch,
model=model,
optimizer=optimizer,
scheduler=scheduler,
metadata={"timestamp": time.time()},
)
self.console.print(f"Saved checkpoint to {path}")
# endregion
def train(self) -> None:
torch.manual_seed(self.training_config.seed)
train_loader, val_loader = self._load_data()
model = self._create_model()
criterion = nn.CrossEntropyLoss()
optimizer = build_optimizer(model, self.training_config)
scheduler = build_scheduler(optimizer, self.training_config)
scaler = torch.cuda.amp.GradScaler() if self.training_config.mixed_precision else None
state = self._restore_checkpoint(model, optimizer, scheduler)
self.global_step = state.get("step", 0)
initial_preview = iter_batch_preview(train_loader, max_batches=1)
if initial_preview:
images, labels = initial_preview[0]
images = images.detach().cpu()
labels = labels.detach().cpu()
self.dashboard.update_preview(
LivePreview(images=images, labels=labels)
)
self.console.print(
f"Starting training for {self.training_config.epochs} epochs on {self.device}"
)
live: Live | None = None
if self.training_config.enable_live:
live = Live(self.dashboard.render(), console=self.console, refresh_per_second=4)
live.start()
try:
for epoch in range(state.get("epoch", 0), self.training_config.epochs):
model.train()
for batch_index, (images, labels) in enumerate(train_loader, start=1):
if (
self.training_config.limit_train_batches
and batch_index > self.training_config.limit_train_batches
):
break
self.global_step += 1
images = images.to(self.device)
labels = labels.to(self.device)
optimizer.zero_grad()
with torch.cuda.amp.autocast(enabled=bool(scaler)):
logits = model(images)
loss = criterion(logits, labels)
preds = logits.detach()
if scaler:
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
else:
loss.backward()
gradient_norm = compute_gradient_norm(model)
if self.training_config.gradient_clip:
torch.nn.utils.clip_grad_norm_(
model.parameters(), self.training_config.gradient_clip
)
if scaler:
scaler.step(optimizer)
scaler.update()
else:
optimizer.step()
if scheduler:
scheduler.step()
accuracy = compute_accuracy(preds, labels)
learning_rate = optimizer.param_groups[0]["lr"]
metrics = self.metrics_tracker.update(
step=self.global_step,
epoch=epoch + 1,
loss=loss,
accuracy=accuracy,
gradient_norm=gradient_norm,
learning_rate=learning_rate,
)
self.dashboard.update_metrics(metrics)
if (
self.training_config.preview_interval
and self.training_config.preview_interval > 0
and self.global_step % self.training_config.preview_interval == 0
):
preview_batches = iter_batch_preview([
(images.detach().cpu(), labels.detach().cpu())
],
max_batches=1,
)
preview_images, preview_labels = preview_batches[0]
predicted_classes = preds.argmax(dim=1).cpu()
self.dashboard.update_preview(
LivePreview(
images=preview_images,
labels=preview_labels,
predictions=predicted_classes,
)
)
if (
self.training_config.log_interval
and self.training_config.log_interval > 0
and self.global_step % self.training_config.log_interval == 0
):
self._write_metrics_log(
{
"phase": "train",
"step": self.global_step,
"epoch": epoch + 1,
"loss": metrics.loss,
"accuracy": metrics.accuracy,
"gradient_norm": metrics.gradient_norm,
"learning_rate": metrics.learning_rate,
}
)
self._maybe_checkpoint(
model=model,
optimizer=optimizer,
scheduler=scheduler,
epoch=epoch + 1,
step=self.global_step,
)
if live:
live.update(self.dashboard.render())
if (epoch + 1) % max(self.training_config.evaluate_every, 1) == 0:
val_metrics = self.evaluate(model, val_loader, epoch + 1)
self._write_metrics_log({"phase": "val", **val_metrics})
self.console.print(
f"Epoch {epoch + 1}: val_loss={val_metrics['loss']:.4f} "
f"val_acc={val_metrics['accuracy']:.4f}"
)
if live:
live.update(self.dashboard.render())
finally:
if live:
live.stop()
self._log_file.close()
def evaluate(
self, model: nn.Module, val_loader: Iterable, epoch: int
) -> Dict[str, float]:
model.eval()
criterion = nn.CrossEntropyLoss()
total_loss = 0.0
total_correct = 0
total_samples = 0
device = self.device
with torch.no_grad():
for batch_index, (images, labels) in enumerate(val_loader, start=1):
if (
self.training_config.limit_val_batches
and batch_index > self.training_config.limit_val_batches
):
break
images = images.to(device)
labels = labels.to(device)
logits = model(images)
loss = criterion(logits, labels)
total_loss += float(loss.item()) * images.size(0)
total_correct += int(logits.argmax(dim=1).eq(labels).sum().item())
total_samples += int(images.size(0))
avg_loss = total_loss / max(total_samples, 1)
accuracy = total_correct / max(total_samples, 1)
metrics = {
"phase": "val",
"epoch": epoch,
"loss": avg_loss,
"accuracy": accuracy,
"step": self.global_step,
}
self.dashboard.log.append(
f"Epoch {epoch} validation: loss {avg_loss:.4f}, accuracy {accuracy:.4f}"
)
return metrics