Repository navigation
Expand file tree
/
Copy pathtrain.py
More file actions
750 lines (653 loc) · 26.8 KB
/
Copy pathtrain.py
File metadata and controls
750 lines (653 loc) · 26.8 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
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
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
import os
import torch
import torch.nn as nn
import torch.optim as optim
import yaml
import argparse
import logging
import numpy as np
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
import wandb
from models.text_jepa import TextJEPA
from data.c4_dataset import create_c4_dataloader
from utils.logger import setup_logger
from utils.metrics import TextJEPAMetrics
def parse_args():
parser = argparse.ArgumentParser(description="Train Text-JEPA model on C4 dataset")
parser.add_argument(
"--config", type=str, default="config/default.yaml", help="Path to config file"
)
parser.add_argument(
"--output_dir", type=str, default="outputs", help="Output directory"
)
parser.add_argument("--log_dir", type=str, default="logs", help="Log directory")
parser.add_argument("--seed", type=int, default=42, help="Random seed")
parser.add_argument(
"--subset",
type=str,
default="en",
help="C4 subset to use (e.g., 'en', 'realnewslike')",
)
parser.add_argument(
"--streaming", action="store_true", help="Use streaming mode for dataset"
)
parser.add_argument(
"--resume",
type=str,
default=None,
help="Path to checkpoint for resuming training",
)
parser.add_argument(
"--debug", action="store_true", help="Enable debug mode with fewer samples"
)
parser.add_argument(
"--experiment_name",
type=str,
default=None,
help="Name for the experiment, used in wandb and outputs",
)
parser.add_argument(
"--use_wandb", action="store_true", help="Enable Weights & Biases logging"
)
parser.add_argument(
"--wandb_project",
type=str,
default="text-jepa",
help="Weights & Biases project name",
)
parser.add_argument(
"--wandb_entity",
type=str,
default=None,
help="Weights & Biases entity (username or team name)",
)
parser.add_argument(
"--wandb_tags",
type=str,
default=None,
help="Comma-separated list of tags for wandb",
)
# Add model initialization method override
parser.add_argument(
"--model_init_method",
type=str,
choices=["pretrained", "from_config", "auto"],
default=None,
help="Override model initialization method from config",
)
# Add new arguments for gradient checkpointing and accumulation
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Enable gradient checkpointing to save memory",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=None,
help="Number of steps to accumulate gradients over",
)
return parser.parse_args()
def load_config(config_path):
with open(config_path, "r") as f:
config = yaml.safe_load(f)
return config
def set_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
def train_with_accumulation(
model, batch, optimizer, config, global_step, train_metrics, device, writer, args
):
"""
Process a single training step with gradient accumulation support
Args:
model: The Text-JEPA model
batch: The current batch of data
optimizer: The optimizer
config: Configuration dictionary
global_step: Current global step counter
train_metrics: Metrics tracker
device: Training device
writer: TensorBoard writer
args: Command line arguments
Returns:
loss: The loss value
metrics_dict: Dictionary of metrics
"""
# Get gradient accumulation steps from config
accumulation_steps = config["training"].get("gradient_accumulation_steps", 1)
# Only zero gradients at the beginning of accumulation cycle
if global_step % accumulation_steps == 0:
optimizer.zero_grad()
# Move batch to device
context_input_ids = batch["context_input_ids"].to(device)
context_attention_mask = batch["context_attention_mask"].to(device)
target_input_ids = batch["target_input_ids"].to(device)
target_attention_mask = batch["target_attention_mask"].to(device)
span_positions = batch["span_positions"].to(device)
# Forward pass
loss, batch_metrics = model(
context_tokens=context_input_ids,
target_tokens=target_input_ids,
span_positions=span_positions,
)
# Scale the loss to account for gradient accumulation
loss = loss / accumulation_steps
# Backward pass
loss.backward()
# Update metrics
train_metrics.update(
batch_metrics["predicted_reprs"],
batch_metrics["target_reprs"],
{
"avg_nonmatching_pred_similarity": batch_metrics[
"avg_nonmatching_pred_similarity"
],
"max_nonmatching_pred_similarity": batch_metrics[
"max_nonmatching_pred_similarity"
],
"min_nonmatching_pred_similarity": batch_metrics[
"min_nonmatching_pred_similarity"
],
"avg_nonmatching_target_similarity": batch_metrics[
"avg_nonmatching_target_similarity"
],
"max_nonmatching_target_similarity": batch_metrics[
"max_nonmatching_target_similarity"
],
"min_nonmatching_target_similarity": batch_metrics[
"min_nonmatching_target_similarity"
],
},
)
# Only update weights at the end of accumulation cycle
if (global_step + 1) % accumulation_steps == 0 or (
global_step + 1 == config["training"]["max_steps"]
):
# Calculate gradient norm for monitoring
total_norm = 0
for p in model.parameters():
if p.grad is not None:
param_norm = p.grad.data.norm(2)
total_norm += param_norm.item() ** 2
total_norm = total_norm**0.5
writer.add_scalar("train/grad_norm", total_norm, global_step)
if args.use_wandb:
wandb.log({"train/grad_norm": total_norm}, step=global_step)
# Clip gradients
torch.nn.utils.clip_grad_norm_(
model.parameters(),
config["training"]["max_grad_norm"],
)
# Update parameters
optimizer.step()
# Compute metrics
metrics_dict = train_metrics.compute()
return loss * accumulation_steps, metrics_dict # Return unscaled loss for logging
def main():
# Parse arguments and load config
args = parse_args()
config = load_config(args.config)
# Set up experiment name
if args.experiment_name is None:
timestamp = torch.cuda.current_device() if torch.cuda.is_available() else ""
args.experiment_name = f"text_jepa_{args.subset}_{timestamp}"
# Set up output directories with experiment name
experiment_output_dir = os.path.join(args.output_dir, args.experiment_name)
experiment_log_dir = os.path.join(args.log_dir, args.experiment_name)
os.makedirs(experiment_output_dir, exist_ok=True)
os.makedirs(experiment_log_dir, exist_ok=True)
# Setup logging
logger = setup_logger("text_jepa", os.path.join(experiment_log_dir, "train.log"))
logger.info(f"Experiment name: {args.experiment_name}")
logger.info(f"Config: {config}")
logger.info(f"Args: {args}")
# Apply model initialization method override if specified
if args.model_init_method is not None:
if "model" not in config:
config["model"] = {}
config["model"]["initialization_method"] = args.model_init_method
logger.info(
f"Overriding model initialization method to: {args.model_init_method}"
)
# Override gradient checkpointing if specified
if args.gradient_checkpointing:
if "model" not in config:
config["model"] = {}
config["model"]["gradient_checkpointing"] = True
logger.info("Gradient checkpointing enabled via command line argument")
# Override gradient accumulation steps if specified
if args.gradient_accumulation_steps is not None:
if "training" not in config:
config["training"] = {}
config["training"][
"gradient_accumulation_steps"
] = args.gradient_accumulation_steps
logger.info(
f"Gradient accumulation steps set to {args.gradient_accumulation_steps}"
)
if "evaluation" not in config:
config["evaluation"] = {}
if "metrics_window_size" not in config["evaluation"]:
config["evaluation"]["metrics_window_size"] = 100 # Default window size for 100 batches
# Get gradient accumulation steps
accumulation_steps = config["training"].get("gradient_accumulation_steps", 1)
# Calculate effective batch size
effective_batch_size = config["training"]["batch_size"] * accumulation_steps
logger.info(
f"Effective batch size: {effective_batch_size} (batch_size: {config['training']['batch_size']} × "
f"accumulation_steps: {accumulation_steps})"
)
# Log memory-saving techniques
memory_saving_techniques = []
if config.get("model", {}).get("gradient_checkpointing", False):
memory_saving_techniques.append("gradient checkpointing")
if accumulation_steps > 1:
memory_saving_techniques.append("gradient accumulation")
if memory_saving_techniques:
logger.info(
f"Memory-saving techniques enabled: {', '.join(memory_saving_techniques)}"
)
# Set random seed
set_seed(args.seed)
# Initialize wandb if enabled
if args.use_wandb:
wandb_tags = args.wandb_tags.split(",") if args.wandb_tags else []
wandb.init(
project=args.wandb_project,
entity=args.wandb_entity,
name=args.experiment_name,
config={**config, **vars(args)},
tags=wandb_tags,
)
logger.info(
f"Weights & Biases initialized with project: {args.wandb_project}, name: {args.experiment_name}"
)
# Setup tensorboard
writer = SummaryWriter(log_dir=experiment_log_dir)
# Setup device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Using device: {device}")
# Create model from config
logger.info("Creating model from config...")
model = TextJEPA.from_config(config)
model = model.to(device)
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
logger.info(f"Model created with {total_params:,} trainable parameters")
# Log model architecture to wandb if enabled
if args.use_wandb:
wandb.watch(model, log="all", log_freq=100)
# Setup metrics
train_metrics = TextJEPAMetrics(
window_size=config["evaluation"]["metrics_window_size"],
use_rolling=True
)
val_metrics = TextJEPAMetrics(use_rolling=False) # Keep reset-based for evaluation
# Load data
logger.info("Loading C4 dataset...")
# Use a larger buffer size for streaming
buffer_size = config["data"]["buffer_size"]
if args.streaming:
buffer_size = max(
buffer_size, 50000
) # Use at least 50k examples in buffer for streaming
# Update config with buffer size
data_config = config.get("data", {})
data_config["buffer_size"] = buffer_size
config["data"] = data_config
train_dataloader = create_c4_dataloader(
config=config,
split="train",
subset=args.subset,
seed=args.seed,
streaming=args.streaming,
)
val_dataloader = create_c4_dataloader(
config=config,
split="validation",
subset=args.subset,
seed=args.seed,
streaming=args.streaming,
)
# Setup optimizer
optimizer = optim.AdamW(
model.parameters(),
lr=config["training"]["learning_rate"],
weight_decay=config["training"]["weight_decay"],
)
# Setup learning rate scheduler
warmup_steps = int(config["training"]["warmup_steps"])
if config["training"]["scheduler"] == "linear_warmup_cosine_decay":
scheduler = optim.lr_scheduler.OneCycleLR(
optimizer,
max_lr=config["training"]["learning_rate"],
total_steps=config["training"]["max_steps"],
pct_start=warmup_steps / config["training"]["max_steps"],
anneal_strategy="cos",
)
elif config["training"]["scheduler"] == "linear_warmup":
def lr_lambda(current_step):
if current_step < warmup_steps:
return float(current_step) / float(max(1, warmup_steps))
return 1.0
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)
else:
scheduler = None
# Resume training if checkpoint provided
start_epoch = 0
global_step = 0
best_val_loss = float("inf")
if args.resume:
logger.info(f"Resuming training from checkpoint: {args.resume}")
checkpoint = torch.load(args.resume, map_location=device)
model.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
if scheduler is not None and "scheduler_state_dict" in checkpoint:
scheduler.load_state_dict(checkpoint["scheduler_state_dict"])
start_epoch = checkpoint["epoch"] + 1
global_step = checkpoint.get("global_step", 0)
best_val_loss = checkpoint.get("val_loss", float("inf"))
logger.info(f"Resumed from epoch {start_epoch}, global step {global_step}")
# Training loop
logger.info("Starting training...")
max_steps = config["training"]["max_steps"]
eval_steps = config["training"]["eval_steps"]
save_steps = config["training"]["save_steps"]
epoch = start_epoch
early_stop = False
# Adjust for debug mode
if args.debug:
logger.info("DEBUG MODE ENABLED: Using fewer steps")
max_steps = min(max_steps, 100)
eval_steps = min(eval_steps, 20)
save_steps = min(save_steps, 50)
config["training"]["eval_samples"] = min(config["training"]["eval_samples"], 64)
# Main training loop - runs until max_steps reached
while global_step < max_steps and not early_stop:
# Reset train metrics for this epoch
train_metrics.reset()
# Training
model.train()
# In streaming mode, we don't have an epoch concept, so we
# just process batches until the next evaluation point
steps_in_epoch = min(
eval_steps,
len(train_dataloader) if not args.streaming else float("inf"),
)
train_progress_bar = tqdm(
enumerate(train_dataloader),
desc=f"Epoch {epoch+1} [Train]",
total=steps_in_epoch,
)
# Process batches until evaluation or end of epoch
steps_since_eval = 0
for step_idx, batch in train_progress_bar:
try:
# Process a training step with gradient accumulation
loss, metrics_dict = train_with_accumulation(
model=model,
batch=batch,
optimizer=optimizer,
config=config,
global_step=global_step,
train_metrics=train_metrics,
device=device,
writer=writer,
args=args,
)
# Update scheduler if we've completed an accumulation cycle
if (global_step + 1) % accumulation_steps == 0 or (
global_step + 1 == max_steps
):
if scheduler is not None:
scheduler.step()
except Exception as e:
logger.error(f"Error during forward/backward pass: {e}")
logger.error(
f"Input shapes: context={batch['context_input_ids'].shape}, "
f"target={batch['target_input_ids'].shape}, spans={batch['span_positions'].shape}"
)
# Skip this batch but continue training
continue
# Update global step
global_step += 1
steps_since_eval += 1
# Update progress bar
train_progress_bar.set_postfix(
{
"loss": f"{loss.item():.4f}",
"cos_sim": f"{metrics_dict['cosine_similarity']:.4f}",
"lr": f"{optimizer.param_groups[0]['lr']:.2e}",
"step": global_step,
}
)
# Log to tensorboard
writer.add_scalar("train/loss", loss.item(), global_step)
writer.add_scalar(
"train/cosine_similarity",
metrics_dict["cosine_similarity"],
global_step,
)
writer.add_scalar(
"train/avg_nonmatching_pred_similarity",
metrics_dict["avg_nonmatching_pred_similarity"],
global_step,
)
writer.add_scalar(
"train/avg_nonmatching_target_similarity",
metrics_dict["avg_nonmatching_target_similarity"],
global_step,
)
writer.add_scalar("train/lr", optimizer.param_groups[0]["lr"], global_step)
# Log to wandb if enabled
if args.use_wandb:
wandb.log(
{
"train/loss": loss.item(),
"train/cosine_similarity": metrics_dict["cosine_similarity"],
"train/avg_nonmatching_pred_similarity": metrics_dict[
"avg_nonmatching_pred_similarity"
],
"train/avg_nonmatching_target_similarity": metrics_dict[
"avg_nonmatching_target_similarity"
],
"train/lr": optimizer.param_groups[0]["lr"],
"epoch": epoch,
"global_step": global_step,
},
step=global_step,
)
# Save checkpoint at regular intervals
if global_step % save_steps == 0:
checkpoint_path = os.path.join(
experiment_output_dir, f"checkpoint-{global_step}.pt"
)
torch.save(
{
"epoch": epoch,
"global_step": global_step,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": (
scheduler.state_dict() if scheduler else None
),
"config": config,
"args": vars(args),
},
checkpoint_path,
)
logger.info(
f"Saved checkpoint at step {global_step} to {checkpoint_path}"
)
# Check if max steps reached
if global_step >= max_steps:
early_stop = True
break
# Check if it's time for evaluation
if steps_since_eval >= eval_steps:
break
# Evaluation
if global_step % eval_steps == 0 or early_stop:
logger.info(f"Evaluating at step {global_step}...")
model.eval()
val_metrics.reset()
# Calculate how many batches to evaluate
eval_batches = (
config["training"]["eval_samples"] // config["training"]["batch_size"]
)
val_progress_bar = tqdm(
enumerate(val_dataloader),
desc=f"Epoch {epoch+1} [Val]",
total=eval_batches,
)
total_val_samples = 0
with torch.no_grad():
for val_step_idx, batch in val_progress_bar:
# Move batch to device
context_input_ids = batch["context_input_ids"].to(device)
context_attention_mask = batch["context_attention_mask"].to(device)
target_input_ids = batch["target_input_ids"].to(device)
target_attention_mask = batch["target_attention_mask"].to(device)
span_positions = batch["span_positions"].to(device)
# Forward pass
loss, batch_metrics = model(
context_tokens=context_input_ids,
target_tokens=target_input_ids,
span_positions=span_positions,
)
# Update metrics
val_metrics.update(
batch_metrics["predicted_reprs"],
batch_metrics["target_reprs"],
{
"avg_nonmatching_pred_similarity": batch_metrics[
"avg_nonmatching_pred_similarity"
],
"max_nonmatching_pred_similarity": batch_metrics[
"max_nonmatching_pred_similarity"
],
"min_nonmatching_pred_similarity": batch_metrics[
"min_nonmatching_pred_similarity"
],
"avg_nonmatching_target_similarity": batch_metrics[
"avg_nonmatching_target_similarity"
],
"max_nonmatching_target_similarity": batch_metrics[
"max_nonmatching_target_similarity"
],
"min_nonmatching_target_similarity": batch_metrics[
"min_nonmatching_target_similarity"
],
},
)
# Increment sample count and update progress bar
total_val_samples += context_input_ids.size(0)
metrics_dict = val_metrics.compute()
val_progress_bar.set_postfix(
{
"loss": f"{loss.item():.4f}",
"cos_sim": f"{metrics_dict['cosine_similarity']:.4f}",
"samples": total_val_samples,
}
)
# Break after evaluating enough samples
if total_val_samples >= config["training"]["eval_samples"]:
break
# Compute validation metrics once at the end
val_metrics_dict = val_metrics.compute()
val_loss = val_metrics_dict["l2_loss"]
val_cosine_sim = val_metrics_dict["cosine_similarity"]
# Log validation metrics
logger.info(
f"Validation at step {global_step} - Loss: {val_loss:.4f}, "
f"Cosine Similarity: {val_cosine_sim:.4f}, "
f"Avg Pred Similarity: {val_metrics_dict['avg_nonmatching_pred_similarity']:.4f}, "
f"Avg Target Similarity: {val_metrics_dict['avg_nonmatching_target_similarity']:.4f}"
)
writer.add_scalar("val/loss", val_loss, global_step)
writer.add_scalar("val/cosine_similarity", val_cosine_sim, global_step)
writer.add_scalar(
"val/avg_nonmatching_pred_similarity",
val_metrics_dict["avg_nonmatching_pred_similarity"],
global_step,
)
writer.add_scalar(
"val/avg_nonmatching_target_similarity",
val_metrics_dict["avg_nonmatching_target_similarity"],
global_step,
)
# Log to wandb if enabled
if args.use_wandb:
wandb.log(
{
"val/loss": val_loss,
"val/cosine_similarity": val_cosine_sim,
"val/avg_nonmatching_pred_similarity": val_metrics_dict[
"avg_nonmatching_pred_similarity"
],
"val/avg_nonmatching_target_similarity": val_metrics_dict[
"avg_nonmatching_target_similarity"
],
},
step=global_step,
)
# Save checkpoint if validation loss improves
if val_loss < best_val_loss:
best_val_loss = val_loss
# Save model checkpoint
checkpoint_path = os.path.join(experiment_output_dir, "best_model.pt")
torch.save(
{
"epoch": epoch,
"global_step": global_step,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": (
scheduler.state_dict() if scheduler else None
),
"val_loss": best_val_loss,
"config": config,
"args": vars(args),
},
checkpoint_path,
)
logger.info(
f"New best model saved at {checkpoint_path} with validation loss: {best_val_loss:.4f}"
)
if args.use_wandb:
wandb.run.summary["best_val_loss"] = best_val_loss
wandb.run.summary["best_val_step"] = global_step
# Move to next epoch
epoch += 1
# Save final model
final_checkpoint_path = os.path.join(experiment_output_dir, "final_model.pt")
torch.save(
{
"epoch": epoch,
"global_step": global_step,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": scheduler.state_dict() if scheduler else None,
"val_loss": best_val_loss,
"config": config,
"args": vars(args),
},
final_checkpoint_path,
)
# Final logging
logger.info("Training completed!")
logger.info(f"Final model saved at {final_checkpoint_path}")
logger.info(f"Best validation loss: {best_val_loss:.4f}")
logger.info(f"Total training steps: {global_step}")
# Close tensorboard writer
writer.close()
# Finish wandb run if enabled
if args.use_wandb:
wandb.run.summary["final_val_loss"] = best_val_loss
wandb.run.summary["total_steps"] = global_step
wandb.run.summary["total_epochs"] = epoch
wandb.finish()
if __name__ == "__main__":
main()