File size: 16,247 Bytes
1f39ae1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
ULTRA-FAST training script for morphological reinflection using aggressive optimizations
"""

import argparse
import json
import logging
import os
import time
import gc
from pathlib import Path
from typing import Dict, Tuple, Optional

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.cuda.amp import GradScaler, autocast
import torch.backends.cudnn as cudnn

from transformer import TagTransformer, PAD_IDX, DEVICE
from morphological_dataset import MorphologicalDataset, build_vocabulary, collate_fn

# Aggressive optimizations
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.allow_tf32 = True
torch.backends.cuda.matmul.allow_tf32 = True

# Minimal logging for speed
logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger(__name__)

def create_optimized_model(config: Dict, src_vocab: Dict[str, int], tgt_vocab: Dict[str, int]) -> TagTransformer:
    """Create and initialize the TagTransformer model with aggressive optimizations"""
    
    # Count feature tokens
    feature_tokens = [token for token in src_vocab.keys() 
                     if token.startswith('<') and token.endswith('>')]
    nb_attr = len(feature_tokens)
    
    model = TagTransformer(
        src_vocab_size=len(src_vocab),
        trg_vocab_size=len(tgt_vocab),
        embed_dim=config['embed_dim'],
        nb_heads=config['nb_heads'],
        src_hid_size=config['src_hid_size'],
        src_nb_layers=config['src_nb_layers'],
        trg_hid_size=config['trg_hid_size'],
        trg_nb_layers=config['trg_nb_layers'],
        dropout_p=config['dropout_p'],
        tie_trg_embed=config['tie_trg_embed'],
        label_smooth=config['label_smooth'],
        nb_attr=nb_attr,
        src_c2i=src_vocab,
        trg_c2i=tgt_vocab,
        attr_c2i={},
    )
    
    # Aggressive weight initialization
    for p in model.parameters():
        if p.dim() > 1:
            nn.init.xavier_uniform_(p)
        elif p.dim() == 1:
            nn.init.uniform_(p, -0.1, 0.1)
    
    # Compile model for speed (PyTorch 2.0+)
    if hasattr(torch, 'compile'):
        try:
            model = torch.compile(model, mode="max-autotune")
            print("✓ Model compiled with torch.compile")
        except Exception as e:
            print(f"⚠ torch.compile failed: {e}")
    
    return model

def create_ultra_fast_dataloader(dataset, config: Dict, src_vocab: Dict, tgt_vocab: Dict):
    """Create ultra-fast DataLoader with minimal overhead"""
    
    # Calculate optimal workers based on system
    num_workers = min(16, os.cpu_count() or 1)
    
    # Use shared memory for faster inter-process communication
    dataloader = DataLoader(
        dataset, 
        batch_size=config['batch_size'], 
        shuffle=True, 
        collate_fn=lambda batch: collate_fn(batch, src_vocab, tgt_vocab, config['max_length']),
        num_workers=num_workers,
        pin_memory=True,
        persistent_workers=True,
        prefetch_factor=4,  # Increased prefetching
        drop_last=True,
        generator=torch.Generator(device='cpu'),  # Deterministic shuffling
    )
    
    return dataloader

def train_epoch_ultra_fast(model: TagTransformer, 
                          dataloader: DataLoader, 
                          optimizer: optim.Optimizer,
                          device: torch.device,
                          epoch: int,
                          config: Dict,
                          scaler: GradScaler) -> float:
    """Ultra-fast training with minimal overhead"""
    
    model.train()
    total_loss = 0.0
    num_batches = 0
    
    # Pre-allocate tensors to avoid repeated allocation
    accumulation_steps = config.get('gradient_accumulation_steps', 1)
    optimizer.zero_grad(set_to_none=True)  # Faster than zero_grad()
    
    # Disable gradient computation for validation tensors
    torch.set_grad_enabled(True)
    
    start_time = time.time()
    
    for batch_idx, (src, src_mask, tgt, tgt_mask) in enumerate(dataloader):
        # Move to device with minimal overhead
        src = src.to(device, non_blocking=True, memory_format=torch.channels_last)
        src_mask = src_mask.to(device, non_blocking=True)
        tgt = tgt.to(device, non_blocking=True, memory_format=torch.channels_last)
        tgt_mask = tgt_mask.to(device, non_blocking=True)
        
        # Mixed precision forward pass
        with autocast(enabled=config.get('use_amp', True)):
            # Forward pass
            output = model(src, src_mask, tgt, tgt_mask)
            
            # Compute loss (shift sequences for next-token prediction)
            loss = model.loss(output[:-1], tgt[1:])
            loss = loss / accumulation_steps
        
        # Mixed precision backward pass
        scaler.scale(loss).backward()
        
        # Gradient accumulation
        if (batch_idx + 1) % accumulation_steps == 0:
            # Gradient clipping
            scaler.unscale_(optimizer)
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=config['gradient_clip'])
            
            # Optimizer step
            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad(set_to_none=True)
        
        total_loss += loss.item() * accumulation_steps
        num_batches += 1
        
        # Minimal logging - only every 100 batches
        if batch_idx % 100 == 0:
            elapsed = time.time() - start_time
            samples_per_sec = (batch_idx + 1) * config['batch_size'] / elapsed
            print(f'Epoch {epoch}, Batch {batch_idx}, Loss: {loss.item() * accumulation_steps:.4f}, Speed: {samples_per_sec:.0f} samples/sec')
    
    # Handle remaining gradients
    if num_batches % accumulation_steps != 0:
        scaler.unscale_(optimizer)
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=config['gradient_clip'])
        scaler.step(optimizer)
        scaler.update()
        optimizer.zero_grad(set_to_none=True)
    
    avg_loss = total_loss / num_batches
    return avg_loss

def validate_fast(model: TagTransformer, 
                 dataloader: DataLoader, 
                 device: torch.device,
                 config: Dict) -> float:
    """Fast validation with minimal overhead"""
    
    model.eval()
    total_loss = 0.0
    num_batches = 0
    
    with torch.no_grad():
        for src, src_mask, tgt, tgt_mask in dataloader:
            src = src.to(device, non_blocking=True, memory_format=torch.channels_last)
            src_mask = src_mask.to(device, non_blocking=True)
            tgt = tgt.to(device, non_blocking=True, memory_format=torch.channels_last)
            tgt_mask = tgt_mask.to(device, non_blocking=True)
            
            with autocast(enabled=config.get('use_amp', True)):
                output = model(src, src_mask, tgt, tgt_mask)
                loss = model.loss(output[:-1], tgt[1:])
            
            total_loss += loss.item()
            num_batches += 1
    
    avg_loss = total_loss / num_batches
    return avg_loss

def save_checkpoint_fast(model: TagTransformer, 
                        optimizer: optim.Optimizer, 
                        epoch: int, 
                        loss: float, 
                        save_path: str,
                        scaler: GradScaler = None):
    """Fast checkpoint saving"""
    
    # Save only essential data
    checkpoint = {
        'epoch': epoch,
        'model_state_dict': model.state_dict(),
        'optimizer_state_dict': optimizer.state_dict(),
        'loss': loss,
    }
    
    if scaler is not None:
        checkpoint['scaler_state_dict'] = scaler.state_dict()
    
    # Use torch.save with _use_new_zipfile_serialization=False for speed
    torch.save(checkpoint, save_path, _use_new_zipfile_serialization=False)
    print(f'Checkpoint saved: {save_path}')

def load_checkpoint_fast(model: TagTransformer, 
                        optimizer: optim.Optimizer, 
                        checkpoint_path: str,
                        scaler: GradScaler = None) -> int:
    """Fast checkpoint loading"""
    
    checkpoint = torch.load(checkpoint_path, map_location=DEVICE, weights_only=False)
    model.load_state_dict(checkpoint['model_state_dict'])
    optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
    
    if scaler is not None and 'scaler_state_dict' in checkpoint:
        scaler.load_state_dict(checkpoint['scaler_state_dict'])
    
    epoch = checkpoint['epoch']
    loss = checkpoint['loss']
    print(f'Checkpoint loaded: {checkpoint_path}, Epoch: {epoch}, Loss: {loss:.4f}')
    return epoch

def main():
    parser = argparse.ArgumentParser(description='ULTRA-FAST training for morphological reinflection')
    parser.add_argument('--resume', type=str, help='Path to checkpoint to resume from')
    parser.add_argument('--output_dir', type=str, default='./models', help='Output directory')
    parser.add_argument('--no_amp', action='store_true', help='Disable mixed precision training')
    parser.add_argument('--profile', action='store_true', help='Enable profiling for debugging')
    args = parser.parse_args()
    
    # Ultra-aggressive configuration for maximum speed
    config = {
        'embed_dim': 256,
        'nb_heads': 4,
        'src_hid_size': 1024,
        'src_nb_layers': 4,
        'trg_hid_size': 1024,
        'trg_nb_layers': 4,
        'dropout_p': 0.1,
        'tie_trg_embed': True,
        'label_smooth': 0.1,
        'batch_size': 800,  # Larger batch size for better GPU utilization
        'learning_rate': 0.001,
        'max_epochs': 1000,
        'max_updates': 10000,
        'warmup_steps': 4000,
        'weight_decay': 0.01,
        'gradient_clip': 1.0,
        'save_every': 20,  # Save less frequently for speed
        'eval_every': 10,  # Evaluate less frequently for speed
        'max_length': 100,
        'use_amp': not args.no_amp,
        'gradient_accumulation_steps': 1,  # No accumulation for maximum speed
        'pin_memory': True,
        'persistent_workers': True,
        'prefetch_factor': 4,
    }
    
    # Create output directory
    os.makedirs(args.output_dir, exist_ok=True)
    os.makedirs(os.path.join(args.output_dir, 'checkpoints'), exist_ok=True)
    
    # Save config
    with open(os.path.join(args.output_dir, 'config.json'), 'w') as f:
        json.dump(config, f, indent=2)
    
    # Set device and enable optimizations
    device = DEVICE
    print(f'Using device: {device}')
    
    # Aggressive CUDA optimizations
    if torch.cuda.is_available():
        # Set memory fraction to avoid fragmentation
        torch.cuda.set_per_process_memory_fraction(0.95)
        
        # Enable memory pool for faster allocation
        torch.cuda.empty_cache()
        gc.collect()
        
        print("✓ CUDA optimizations enabled")
        print(f"GPU: {torch.cuda.get_device_name()}")
        print(f"Memory: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.1f} GB")
    
    # Data file paths
    train_src = '../10L_90NL/train/run1/train.10L_90NL_1_1.src'
    train_tgt = '../10L_90NL/train/run1/train.10L_90NL_1_1.tgt'
    dev_src = '../10L_90NL/dev/run1/dev.10L_90NL_1_1.src'
    dev_tgt = '../10L_90NL/dev/run1/dev.10L_90NL_1_1.tgt'
    
    # Build vocabulary efficiently
    print("Building vocabulary...")
    src_vocab = build_vocabulary([train_src, dev_src])
    tgt_vocab = build_vocabulary([train_tgt, dev_tgt])
    
    print(f"Source vocabulary size: {len(src_vocab)}")
    print(f"Target vocabulary size: {len(tgt_vocab)}")
    
    # Create datasets
    train_dataset = MorphologicalDataset(train_src, train_tgt, src_vocab, tgt_vocab, config['max_length'])
    dev_dataset = MorphologicalDataset(dev_src, dev_tgt, src_vocab, tgt_vocab, config['max_length'])
    
    # Create ultra-fast dataloaders
    train_loader = create_ultra_fast_dataloader(train_dataset, config, src_vocab, tgt_vocab)
    dev_loader = create_ultra_fast_dataloader(dev_dataset, config, src_vocab, tgt_vocab)
    
    # Create optimized model
    model = create_optimized_model(config, src_vocab, tgt_vocab)
    model = model.to(device)
    
    # Use channels_last memory format for better performance
    if hasattr(model, 'to'):
        model = model.to(memory_format=torch.channels_last)
    
    # Count parameters
    total_params = model.count_nb_params()
    print(f'Total parameters: {total_params:,}')
    
    # Create optimizer with aggressive settings
    optimizer = optim.AdamW(
        model.parameters(),
        lr=config['learning_rate'],
        weight_decay=config['weight_decay'],
        betas=(0.9, 0.999),
        eps=1e-8,
        foreach=True,  # Use foreach implementation for speed
    )
    
    # Learning rate scheduler
    def lr_lambda(step):
        if step < config['warmup_steps']:
            return float(step) / float(max(1, config['warmup_steps']))
        progress = (step - config['warmup_steps']) / (config['max_updates'] - config['warmup_steps'])
        return max(0.0, 0.5 * (1.0 + torch.cos(torch.pi * progress)))
    
    scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)
    
    # Mixed precision training
    scaler = GradScaler(enabled=config['use_amp'])
    if config['use_amp']:
        print("✓ Mixed precision training enabled")
    
    # Resume from checkpoint if specified
    start_epoch = 0
    if args.resume:
        start_epoch = load_checkpoint_fast(model, optimizer, args.resume, scaler)
    
    # Training loop
    best_val_loss = float('inf')
    global_step = 0
    updates = 0
    
    print(f"\nStarting training with {len(train_loader)} batches per epoch")
    print(f"Effective batch size: {config['batch_size'] * config['gradient_accumulation_steps']}")
    
    for epoch in range(start_epoch, config['max_epochs']):
        epoch_start_time = time.time()
        
        # Train
        train_loss = train_epoch_ultra_fast(
            model, train_loader, optimizer, device, epoch, config, scaler
        )
        
        # Update learning rate
        scheduler.step()
        current_lr = scheduler.get_last_lr()[0]
        
        # Validate less frequently for speed
        if epoch % config['eval_every'] == 0:
            val_loss = validate_fast(model, dev_loader, device, config)
            
            print(f'Epoch {epoch}: Train Loss: {train_loss:.4f}, Val Loss: {val_loss:.4f}, LR: {current_lr:.6f}')
            
            # Save best model
            if val_loss < best_val_loss:
                best_val_loss = val_loss
                save_checkpoint_fast(
                    model, optimizer, epoch, val_loss,
                    os.path.join(args.output_dir, 'checkpoints', 'best_model.pth'),
                    scaler
                )
        else:
            print(f'Epoch {epoch}: Train Loss: {train_loss:.4f}, LR: {current_lr:.6f}')
        
        # Save checkpoint less frequently for speed
        if epoch % config['save_every'] == 0:
            save_checkpoint_fast(
                model, optimizer, epoch, train_loss,
                os.path.join(args.output_dir, 'checkpoints', f'checkpoint_epoch_{epoch}.pth'),
                scaler
            )
        
        epoch_time = time.time() - epoch_start_time
        samples_per_sec = len(train_loader) * config['batch_size'] / epoch_time
        
        print(f'Epoch {epoch} completed in {epoch_time:.2f}s ({samples_per_sec:.0f} samples/sec)')
        
        # Count updates
        updates += len(train_loader)
        global_step += len(train_loader)
        
        # Check if we've reached max updates
        if updates >= config['max_updates']:
            print(f'Reached maximum updates ({config["max_updates"]}), stopping training')
            break
        
        # Clear cache periodically
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
    
    # Save final model
    save_checkpoint_fast(
        model, optimizer, epoch, train_loss,
        os.path.join(args.output_dir, 'checkpoints', 'final_model.pth'),
        scaler
    )
    
    print('Training completed!')

if __name__ == '__main__':
    main()