File size: 10,465 Bytes
c1fcc58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader, random_split
from torch.cuda.amp import autocast, GradScaler
from torch.utils.tensorboard import SummaryWriter
import numpy as np
from tqdm import tqdm
import json
import argparse
from datetime import datetime

# Configuration with Transformer-specific parameters
CONFIG = {
    "FILE_PATH": 'dataset.txt',
    "SEQ_LENGTH": 32,
    "BATCH_SIZE": 8,
    "EPOCHS": 1,
    "EMBEDDING_DIM": 64,
    "N_HEADS": 1,
    "FFN_DIM": 64,
    "NUM_LAYERS": 3,
    "DROPOUT": 0.1,
    "LEARNING_RATE": 0.0005,
    "WEIGHT_DECAY": 0.01,
    "CLIP_GRAD": 1.0,
    "LABEL_SMOOTHING": 0.1,
    "GRAD_ACCUM_STEPS": 2,
    "VAL_SPLIT": 0.1,
    "EARLY_STOP_PATIENCE": 3,
    "MODEL_SAVE_PATH": "transformer_lm_model.pth",
    "TEMPERATURE": 0.7,
    "TOP_K": 50,
    "TOP_P": 0.9,
    "LOG_DIR": "runs"
}

# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
scaler = GradScaler(enabled=device.type == 'cuda')

# Handle command-line arguments
parser = argparse.ArgumentParser()
parser.add_argument('--config', type=str, help='Path to config JSON file')
args = parser.parse_args()

if args.config:
    with open(args.config) as f:
        CONFIG.update(json.load(f))

# Initialize TensorBoard
writer = SummaryWriter(f"{CONFIG['LOG_DIR']}/{datetime.now().strftime('%Y%m%d-%H%M%S')}")

# Read and process text
with open(CONFIG["FILE_PATH"], 'r', encoding='utf-8') as f:
    text = f.read()

# Vocabulary setup
chars = sorted(list(set(text)))
vocab_size = len(chars)
char_to_idx = {ch: i for i, ch in enumerate(chars)}
idx_to_char = {i: ch for i, ch in enumerate(chars)}

# Encode text
encoded_text = np.array([char_to_idx[ch] for ch in text])

# Dataset class with memory mapping
class TextDataset(Dataset):
    def __init__(self, data, seq_length):
        self.data = torch.from_numpy(data).long()
        self.seq_length = seq_length
        
    def __len__(self):
        return len(self.data) - self.seq_length - 1
    
    def __getitem__(self, idx):
        x = self.data[idx:idx+self.seq_length]
        y = self.data[idx+1:idx+self.seq_length+1]
        return x, y

dataset = TextDataset(encoded_text, CONFIG["SEQ_LENGTH"])
val_size = int(len(dataset) * CONFIG["VAL_SPLIT"])
train_size = len(dataset) - val_size
train_dataset, val_dataset = random_split(dataset, [train_size, val_size])

train_loader = DataLoader(train_dataset, batch_size=CONFIG["BATCH_SIZE"], 
                         shuffle=True, pin_memory=True, num_workers=4)
val_loader = DataLoader(val_dataset, batch_size=CONFIG["BATCH_SIZE"], 
                       pin_memory=True, num_workers=4)

# Transformer-based Language Model
class TransformerLM(nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, CONFIG["EMBEDDING_DIM"])
        self.pos_embed = nn.Embedding(CONFIG["SEQ_LENGTH"], CONFIG["EMBEDDING_DIM"])
        
        self.transformer = nn.TransformerEncoder(
            nn.TransformerEncoderLayer(
                d_model=CONFIG["EMBEDDING_DIM"],
                nhead=CONFIG["N_HEADS"],
                dim_feedforward=CONFIG["FFN_DIM"],
                dropout=CONFIG["DROPOUT"],
                activation='gelu',
                batch_first=True
            ),
            num_layers=CONFIG["NUM_LAYERS"]
        )
        
        self.ln = nn.LayerNorm(CONFIG["EMBEDDING_DIM"])
        self.fc = nn.Linear(CONFIG["EMBEDDING_DIM"], vocab_size)
        
        self.apply(self._init_weights)
        
    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            nn.init.xavier_uniform_(module.weight)
            if module.bias is not None:
                nn.init.constant_(module.bias, 0)
        elif isinstance(module, nn.Embedding):
            nn.init.xavier_uniform_(module.weight)
            
    def forward(self, x, mask=None):
        batch_size, seq_len = x.size()
        positions = torch.arange(seq_len, device=device).expand(batch_size, seq_len)
        x = self.embedding(x) + self.pos_embed(positions)
        
        if mask is None:
            mask = nn.Transformer.generate_square_subsequent_mask(seq_len).to(device)
            
        x = self.transformer(x, mask)
        x = self.ln(x)
        return self.fc(x), None  # Return None for compatibility with generation code

model = TransformerLM().to(device)
optimizer = torch.optim.AdamW(model.parameters(), 
                             lr=CONFIG["LEARNING_RATE"], 
                             weight_decay=CONFIG["WEIGHT_DECAY"])
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
    optimizer, mode='min', factor=0.5, patience=2)

# Training loop with advanced features
best_val_loss = float('inf')
patience_counter = 0

for epoch in range(CONFIG["EPOCHS"]):
    model.train()
    train_loss = 0
    progress_bar = tqdm(train_loader, desc=f'Epoch {epoch+1}/{CONFIG["EPOCHS"]}')
    
    for i, (inputs, targets) in enumerate(progress_bar):
        inputs, targets = inputs.to(device), targets.to(device)
        
        with autocast(enabled=device.type == 'cuda'):
            outputs, _ = model(inputs)
            logits = outputs.view(-1, vocab_size)
            targets = targets.view(-1)
            
            if CONFIG["LABEL_SMOOTHING"]:
                loss = F.cross_entropy(logits, targets, 
                                      label_smoothing=CONFIG["LABEL_SMOOTHING"])
            else:
                loss = F.cross_entropy(logits, targets)
            
            loss = loss / CONFIG["GRAD_ACCUM_STEPS"]
        
        scaler.scale(loss).backward()
        
        if (i + 1) % CONFIG["GRAD_ACCUM_STEPS"] == 0:
            scaler.unscale_(optimizer)
            nn.utils.clip_grad_norm_(model.parameters(), CONFIG["CLIP_GRAD"])
            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad()
            
            lr = optimizer.param_groups[0]['lr']
            progress_bar.set_postfix({'loss': loss.item() * CONFIG["GRAD_ACCUM_STEPS"], 'lr': lr})
        
        train_loss += loss.item() * CONFIG["GRAD_ACCUM_STEPS"]
    
    # Validation phase
    model.eval()
    val_loss = 0
    with torch.no_grad():
        for inputs, targets in val_loader:
            inputs, targets = inputs.to(device), targets.to(device)
            outputs, _ = model(inputs)
            loss = F.cross_entropy(outputs.view(-1, vocab_size), targets.view(-1))
            val_loss += loss.item()
    
    avg_train_loss = train_loss / len(train_loader)
    avg_val_loss = val_loss / len(val_loader)
    scheduler.step(avg_val_loss)
    
    # Log metrics
    writer.add_scalar('Loss/train', avg_train_loss, epoch)
    writer.add_scalar('Loss/val', avg_val_loss, epoch)
    writer.add_scalar('Learning Rate', optimizer.param_groups[0]['lr'], epoch)
    writer.add_scalar('Perplexity/train', np.exp(avg_train_loss), epoch)
    writer.add_scalar('Perplexity/val', np.exp(avg_val_loss), epoch)
    
    print(f'Epoch {epoch+1} | Train Loss: {avg_train_loss:.4f} | Val Loss: {avg_val_loss:.4f}')
    
    # Early stopping and checkpointing
    if avg_val_loss < best_val_loss:
        best_val_loss = avg_val_loss
        torch.save({
            'epoch': epoch,
            'model_state_dict': model.state_dict(),
            'optimizer_state_dict': optimizer.state_dict(),
            'scheduler_state_dict': scheduler.state_dict(),
            'config': CONFIG
        }, CONFIG["MODEL_SAVE_PATH"])
        patience_counter = 0
    else:
        patience_counter += 1
        if patience_counter >= CONFIG["EARLY_STOP_PATIENCE"]:
            print("Early stopping triggered")
            break

writer.close()
print(f'Best model saved to {CONFIG["MODEL_SAVE_PATH"]} with validation loss: {best_val_loss:.4f}')

# Advanced generation with multiple sampling strategies
def generate_text(model, start_str, length=200, temperature=CONFIG["TEMPERATURE"],

                 top_k=CONFIG["TOP_K"], top_p=CONFIG["TOP_P"]):
    model.eval()
    chars = list(start_str)
    input_seq = torch.tensor([char_to_idx[ch] for ch in chars], device=device).unsqueeze(0)
    
    with torch.no_grad():
        for _ in tqdm(range(length), desc="Generating text"):
            mask = nn.Transformer.generate_square_subsequent_mask(input_seq.size(1)).to(device)
            outputs, _ = model(input_seq[:, -CONFIG["SEQ_LENGTH"]:], mask)
            logits = outputs[:, -1] / temperature
            
            # Apply nucleus sampling first
            if top_p > 0:
                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
                cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
                sorted_indices_to_remove = cumulative_probs > top_p
                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                sorted_indices_to_remove[..., 0] = 0
                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
                logits = logits.masked_fill(indices_to_remove, float('-inf'))
            
            # Then apply top-k filtering
            if top_k > 0:
                top_k = min(top_k, logits.size(-1))
                indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
                logits = logits.masked_fill(indices_to_remove, float('-inf'))
            
            probs = F.softmax(logits, dim=-1)
            next_char = torch.multinomial(probs, num_samples=1)
            chars.append(idx_to_char[next_char.item()])
            input_seq = torch.cat([input_seq, next_char], dim=1)
    
    return ''.join(chars)

# Generate examples with different parameters
print("\nConservative sampling:")
print(generate_text(model, "The ", temperature=0.5, top_p=0))

print("\nCreative sampling:")
print(generate_text(model, "Once ", temperature=1.2, top_p=0.9))

print("\nTop-k sampling:")
print(generate_text(model, "In ", top_k=50))

print("\nCombined sampling:")
print(generate_text(model, "Artificial intelligence ", temperature=0.8, top_k=50, top_p=0.9))