File size: 13,597 Bytes
36c78b1 |
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 |
#!/usr/bin/env python3
"""
Full end-to-end BitTransformerLM training run with all optimizations!
Small scale test to validate our enhanced system.
"""
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import numpy as np
import logging
from pathlib import Path
import time
from typing import List, Dict, Any
# Import our enhanced modules
from bit_transformer.model import BitTransformerLM
from bit_transformer.compression import compress_bits_batch, model_output_decompress
from bit_transformer.error_handling import safe_model_forward, setup_error_logging
from bit_transformer.types import BitSequence, TelemetryDict
from enhanced_checkpoint_system import create_checkpoint_manager
# Setup logging
logger = setup_error_logging("INFO")
class SimpleBitDataset(Dataset):
"""Simple dataset of bit sequences for training."""
def __init__(self, num_samples: int = 1000, seq_length: int = 128):
self.num_samples = num_samples
self.seq_length = seq_length
self.data = self._generate_bit_sequences()
def _generate_bit_sequences(self) -> List[torch.Tensor]:
"""Generate diverse bit sequences with different patterns."""
sequences = []
# Pattern 1: Alternating sequences
for i in range(self.num_samples // 4):
pattern = torch.tensor([i % 2] * self.seq_length, dtype=torch.long)
sequences.append(pattern)
# Pattern 2: Random sequences
for i in range(self.num_samples // 4):
pattern = torch.randint(0, 2, (self.seq_length,), dtype=torch.long)
sequences.append(pattern)
# Pattern 3: Structured patterns (runs)
for i in range(self.num_samples // 4):
pattern = []
pos = 0
while pos < self.seq_length:
run_length = min(np.random.randint(1, 20), self.seq_length - pos)
bit_value = np.random.randint(0, 2)
pattern.extend([bit_value] * run_length)
pos += run_length
pattern = torch.tensor(pattern[:self.seq_length], dtype=torch.long)
sequences.append(pattern)
# Pattern 4: Fibonacci-like sequences
remaining = self.num_samples - len(sequences)
for i in range(remaining):
pattern = [0, 1]
while len(pattern) < self.seq_length:
pattern.append(pattern[-1] ^ pattern[-2]) # XOR of last two bits
pattern = torch.tensor(pattern[:self.seq_length], dtype=torch.long)
sequences.append(pattern)
return sequences
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
sequence = self.data[idx]
# For language modeling, input is sequence[:-1], target is sequence[1:]
return sequence[:-1], sequence[1:]
def compute_safety_metrics(predictions: torch.Tensor, targets: torch.Tensor) -> Dict[str, float]:
"""Compute K/C/S safety metrics."""
pred_bits = (predictions > 0.5).float().flatten()
# K metric (Negentropy): Measure of order vs randomness
if len(pred_bits) > 0:
prob_1 = pred_bits.mean().item()
prob_0 = 1 - prob_1
if prob_0 > 0 and prob_1 > 0:
entropy = -prob_0 * np.log2(prob_0) - prob_1 * np.log2(prob_1)
negentropy = 1.0 - entropy # Higher = more ordered
else:
negentropy = 1.0 if prob_1 == 1.0 or prob_1 == 0.0 else 0.0
else:
negentropy = 0.0
# C metric (Complexity): Simple run-length approximation
changes = (pred_bits[1:] != pred_bits[:-1]).sum().item()
complexity = min(changes / len(pred_bits), 1.0) if len(pred_bits) > 1 else 0.0
# S metric (Symbiosis): Alignment with target distribution
target_bits = targets.float().flatten()
if len(target_bits) > 0:
target_mean = target_bits.mean()
pred_mean = pred_bits.mean()
symbiosis = 1.0 - abs(target_mean - pred_mean).item()
else:
symbiosis = 1.0
return {
'K_negentropy': negentropy,
'C_complexity': complexity,
'S_symbiosis': symbiosis
}
def train_bittransformer():
"""Main training function with all optimizations."""
logger.info("π Starting BitTransformerLM end-to-end training run!")
# Model configuration - small but meaningful
model_config = {
'd_model': 256,
'nhead': 8,
'num_layers': 4,
'dim_feedforward': 512,
'max_seq_len': 128,
'use_checkpoint': True,
'chunk_size': None, # Disable chunking for small model
}
training_config = {
'batch_size': 16,
'learning_rate': 1e-3,
'num_epochs': 10,
'save_every_n_epochs': 2,
'log_every_n_steps': 10
}
# Initialize enhanced checkpoint manager
checkpoint_manager = create_checkpoint_manager()
session_id = checkpoint_manager.create_training_session(
session_name="end_to_end_test",
model_config=model_config,
training_config=training_config
)
logger.info(f"π Created training session: {session_id}")
# Create dataset and dataloader
logger.info("π Creating training dataset...")
dataset = SimpleBitDataset(num_samples=800, seq_length=model_config['max_seq_len'])
dataloader = DataLoader(dataset, batch_size=training_config['batch_size'], shuffle=True)
# Initialize model
logger.info("π§ Initializing BitTransformerLM model...")
model = BitTransformerLM(
d_model=model_config['d_model'],
nhead=model_config['nhead'],
num_layers=model_config['num_layers'],
dim_feedforward=model_config['dim_feedforward'],
max_seq_len=model_config['max_seq_len'],
use_checkpoint=model_config['use_checkpoint'],
chunk_size=model_config['chunk_size']
)
# Count parameters
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
logger.info(f"π’ Model parameters: {total_params:,} total, {trainable_params:,} trainable")
# Setup optimizer and loss
optimizer = optim.AdamW(model.parameters(), lr=training_config['learning_rate'])
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=training_config['num_epochs'])
criterion = nn.CrossEntropyLoss()
# Training loop
logger.info("πββοΈ Starting training loop...")
for epoch in range(training_config['num_epochs']):
model.train()
epoch_loss = 0.0
epoch_metrics = {'K_negentropy': 0.0, 'C_complexity': 0.0, 'S_symbiosis': 0.0}
num_batches = 0
start_time = time.time()
for batch_idx, (inputs, targets) in enumerate(dataloader):
optimizer.zero_grad()
# Forward pass with safety monitoring
try:
# BitTransformerLM returns (logits, telemetry)
output = safe_model_forward(model, inputs)
if isinstance(output, tuple):
logits, telemetry = output
else:
logits = output
telemetry = {}
# BitTransformerLM outputs logits for binary classification
# Shape should be [batch, seq_len, 2] for binary vocab
if logits.dim() == 2:
# If [batch*seq_len, 2], already flattened
logits_flat = logits
targets_flat = targets.reshape(-1)
else:
# If [batch, seq_len, 2], flatten
logits_flat = logits.reshape(-1, 2)
targets_flat = targets.reshape(-1)
loss = criterion(logits_flat, targets_flat)
# Backward pass
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
# Compute metrics
with torch.no_grad():
# Handle different logits shapes for predictions
if logits.dim() == 2:
# [batch*seq_len, 2] -> reshape back to [batch, seq_len, 2]
batch_size = inputs.shape[0]
seq_len = inputs.shape[1]
logits_reshaped = logits.reshape(batch_size, seq_len, 2)
predictions = torch.softmax(logits_reshaped, dim=-1)[:, :, 1] # Prob of bit=1
else:
# [batch, seq_len, 2]
predictions = torch.softmax(logits, dim=-1)[:, :, 1] # Prob of bit=1
safety_metrics = compute_safety_metrics(predictions, targets)
epoch_loss += loss.item()
for key, value in safety_metrics.items():
epoch_metrics[key] += value
num_batches += 1
# Logging
if batch_idx % training_config['log_every_n_steps'] == 0:
logger.info(f"Epoch {epoch+1}/{training_config['num_epochs']}, "
f"Batch {batch_idx}/{len(dataloader)}, "
f"Loss: {loss.item():.4f}, "
f"K: {safety_metrics['K_negentropy']:.3f}, "
f"C: {safety_metrics['C_complexity']:.3f}, "
f"S: {safety_metrics['S_symbiosis']:.3f}")
except Exception as e:
logger.error(f"Error in batch {batch_idx}: {e}")
continue
# End of epoch processing
scheduler.step()
epoch_time = time.time() - start_time
if num_batches > 0:
avg_loss = epoch_loss / num_batches
avg_metrics = {k: v / num_batches for k, v in epoch_metrics.items()}
logger.info(f"β
Epoch {epoch+1} completed in {epoch_time:.2f}s")
logger.info(f"π Avg Loss: {avg_loss:.4f}")
logger.info(f"π Safety Metrics - K: {avg_metrics['K_negentropy']:.3f}, "
f"C: {avg_metrics['C_complexity']:.3f}, "
f"S: {avg_metrics['S_symbiosis']:.3f}")
# Save checkpoint
if (epoch + 1) % training_config['save_every_n_epochs'] == 0:
checkpoint_success = checkpoint_manager.save_checkpoint(
model=model,
session_id=session_id,
epoch=epoch + 1,
metrics={
'loss': avg_loss,
'learning_rate': scheduler.get_last_lr()[0],
**avg_metrics
},
optimizer_state=optimizer.state_dict(),
scheduler_state=scheduler.state_dict()
)
if checkpoint_success:
logger.info(f"πΎ Checkpoint saved for epoch {epoch+1}")
# Save best model if loss improved
checkpoint_manager.save_best_model(
session_id=session_id,
model=model,
metric_name='loss',
metric_value=avg_loss,
is_better_func=lambda x, y: x < y # Lower loss is better
)
logger.info("π Training completed successfully!")
# Test inference and compression
logger.info("π§ͺ Testing model inference and compression...")
model.eval()
with torch.no_grad():
# Create a test sequence
test_input = torch.randint(0, 2, (1, 64), dtype=torch.long)
logger.info(f"π₯ Input sequence: {test_input.squeeze().tolist()}")
# Model inference
output_logits = model(test_input)
output_probs = torch.softmax(output_logits, dim=-1)
predicted_bits = torch.argmax(output_probs, dim=-1)
logger.info(f"π€ Predicted sequence: {predicted_bits.squeeze().tolist()}")
# Test compression
compressed = compress_bits_batch(predicted_bits)
logger.info(f"ποΈ Compressed length: {len(compressed[0])} (original: {predicted_bits.shape[-1]})")
# Decompress to verify
decompressed = model_output_decompress(compressed)
compression_success = torch.equal(predicted_bits, decompressed)
logger.info(f"β
Compression/decompression successful: {compression_success}")
# Final storage usage report
storage_usage = checkpoint_manager.get_storage_usage()
logger.info(f"πΎ Final storage usage: {storage_usage['total_gb']:.3f} GB")
logger.info(f"π Training sessions: {storage_usage['num_sessions']}")
return session_id, model, checkpoint_manager
if __name__ == "__main__":
try:
session_id, trained_model, manager = train_bittransformer()
print(f"\nπ SUCCESS! Training session completed: {session_id}")
print(f"π Use checkpoint_manager.load_checkpoint('{session_id}') to resume")
except Exception as e:
logger.error(f"β Training failed: {e}")
import traceback
traceback.print_exc()
raise |