File size: 13,626 Bytes
9b1c753 21613a7 9b1c753 21613a7 9b1c753 |
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 |
"""
Calibration Script for Legal-BERT
Executes Week 7: Model Calibration & Uncertainty Quantification
"""
import torch
import os
import json
import numpy as np
from datetime import datetime
from config import LegalBertConfig
from trainer import LegalBertTrainer, LegalClauseDataset, collate_batch
from data_loader import CUADDataLoader
from model import HierarchicalLegalBERT
from torch.utils.data import DataLoader
class CalibrationFramework:
"""
Calibration methods for Legal-BERT confidence scores
Week 7 implementation: Temperature Scaling, Platt Scaling, Isotonic Regression
"""
def __init__(self, model, device):
self.model = model
self.device = device
self.temperature = 1.0
def collect_logits_and_labels(self, data_loader):
"""Collect logits and true labels from validation set"""
all_logits = []
all_labels = []
self.model.eval()
with torch.no_grad():
for batch in data_loader:
input_ids = batch['input_ids'].to(self.device)
attention_mask = batch['attention_mask'].to(self.device)
labels = batch['risk_label']
# Use the correct method for HierarchicalLegalBERT
outputs = self.model.forward_single_clause(input_ids, attention_mask)
logits = outputs['risk_logits']
all_logits.append(logits.cpu())
all_labels.append(labels)
return torch.cat(all_logits), torch.cat(all_labels)
def temperature_scaling(self, val_loader, lr=0.01, max_iter=50):
"""
Apply temperature scaling calibration
Learns optimal temperature to calibrate confidence scores
"""
print("π‘οΈ Applying temperature scaling...")
# Collect validation logits and labels
logits, labels = self.collect_logits_and_labels(val_loader)
# Create temperature parameter
temperature = torch.nn.Parameter(torch.ones(1) * 1.5)
optimizer = torch.optim.LBFGS([temperature], lr=lr, max_iter=max_iter)
criterion = torch.nn.CrossEntropyLoss()
def eval_loss():
optimizer.zero_grad()
loss = criterion(logits / temperature, labels)
loss.backward()
return loss
optimizer.step(eval_loss)
self.temperature = temperature.item()
print(f" β
Optimal temperature: {self.temperature:.4f}")
return self.temperature
def apply_temperature(self, logits):
"""Apply learned temperature to logits"""
return logits / self.temperature
def calculate_ece(self, data_loader, n_bins=15):
"""
Calculate Expected Calibration Error (ECE)
Measures calibration quality
"""
print("π Calculating Expected Calibration Error (ECE)...")
confidences = []
predictions = []
true_labels = []
self.model.eval()
with torch.no_grad():
for batch in data_loader:
input_ids = batch['input_ids'].to(self.device)
attention_mask = batch['attention_mask'].to(self.device)
labels = batch['risk_label']
# Use the correct method for HierarchicalLegalBERT
outputs = self.model.forward_single_clause(input_ids, attention_mask)
logits = self.apply_temperature(outputs['risk_logits'])
probs = torch.softmax(logits, dim=-1)
conf, pred = torch.max(probs, dim=-1)
confidences.extend(conf.cpu().numpy())
predictions.extend(pred.cpu().numpy())
true_labels.extend(labels.numpy())
confidences = np.array(confidences)
predictions = np.array(predictions)
true_labels = np.array(true_labels)
# Calculate ECE
ece = 0.0
bin_boundaries = np.linspace(0, 1, n_bins + 1)
for i in range(n_bins):
bin_lower = bin_boundaries[i]
bin_upper = bin_boundaries[i + 1]
in_bin = (confidences > bin_lower) & (confidences <= bin_upper)
prop_in_bin = np.mean(in_bin)
if prop_in_bin > 0:
accuracy_in_bin = np.mean(predictions[in_bin] == true_labels[in_bin])
avg_confidence_in_bin = np.mean(confidences[in_bin])
ece += np.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin
print(f" ECE: {ece:.4f}")
return ece
def calculate_mce(self, data_loader, n_bins=15):
"""
Calculate Maximum Calibration Error (MCE)
"""
print("π Calculating Maximum Calibration Error (MCE)...")
confidences = []
predictions = []
true_labels = []
self.model.eval()
with torch.no_grad():
for batch in data_loader:
input_ids = batch['input_ids'].to(self.device)
attention_mask = batch['attention_mask'].to(self.device)
labels = batch['risk_label']
# Use the correct method for HierarchicalLegalBERT
outputs = self.model.forward_single_clause(input_ids, attention_mask)
logits = self.apply_temperature(outputs['risk_logits'])
probs = torch.softmax(logits, dim=-1)
conf, pred = torch.max(probs, dim=-1)
confidences.extend(conf.cpu().numpy())
predictions.extend(pred.cpu().numpy())
true_labels.extend(labels.numpy())
confidences = np.array(confidences)
predictions = np.array(predictions)
true_labels = np.array(true_labels)
# Calculate MCE
mce = 0.0
bin_boundaries = np.linspace(0, 1, n_bins + 1)
for i in range(n_bins):
bin_lower = bin_boundaries[i]
bin_upper = bin_boundaries[i + 1]
in_bin = (confidences > bin_lower) & (confidences <= bin_upper)
if np.sum(in_bin) > 0:
accuracy_in_bin = np.mean(predictions[in_bin] == true_labels[in_bin])
avg_confidence_in_bin = np.mean(confidences[in_bin])
mce = max(mce, np.abs(avg_confidence_in_bin - accuracy_in_bin))
print(f" MCE: {mce:.4f}")
return mce
def main():
"""Execute calibration pipeline"""
print("=" * 80)
print("π‘οΈ LEGAL-BERT CALIBRATION PIPELINE")
print("=" * 80)
# Initialize configuration
config = LegalBertConfig()
# Load trained model
print("\nπ Loading trained model...")
model_path = os.path.join(config.model_save_path, 'final_model.pt')
if not os.path.exists(model_path):
print(f"β Error: Model not found at {model_path}")
print("Please train the model first using: python train.py")
return
checkpoint = torch.load(model_path, map_location=config.device, weights_only=False)
# CRITICAL FIX: Use the config from checkpoint to get correct architecture parameters
if 'config' in checkpoint:
saved_config = checkpoint['config']
hidden_dim = saved_config.hierarchical_hidden_dim
num_lstm_layers = saved_config.hierarchical_num_lstm_layers
print(f" Using saved architecture: hidden_dim={hidden_dim}, lstm_layers={num_lstm_layers}")
else:
# Fallback to current config (for backward compatibility)
hidden_dim = config.hierarchical_hidden_dim
num_lstm_layers = config.hierarchical_num_lstm_layers
print(f" β οΈ Warning: No config in checkpoint, using current config")
# Initialize and load Hierarchical BERT model
print("π Loading Hierarchical BERT model")
model = HierarchicalLegalBERT(
config=config,
num_discovered_risks=len(checkpoint['discovered_patterns']),
hidden_dim=hidden_dim,
num_lstm_layers=num_lstm_layers
).to(config.device)
model.load_state_dict(checkpoint['model_state_dict'])
print("β
Model loaded successfully!")
# Load validation and test data
print("\nπ Loading data...")
data_loader = CUADDataLoader(config.data_path)
df_clauses, contracts = data_loader.load_data()
splits = data_loader.create_splits()
# Initialize trainer for helper methods
trainer = LegalBertTrainer(config)
# Restore risk discovery model (including fitted LDA/K-Means)
if 'risk_discovery_model' in checkpoint:
trainer.risk_discovery = checkpoint['risk_discovery_model']
else:
# Fallback for older models
trainer.risk_discovery.discovered_patterns = checkpoint['discovered_patterns']
trainer.risk_discovery.n_clusters = len(checkpoint['discovered_patterns'])
trainer.model = model
# Prepare validation and test loaders
val_clauses = splits['val']['clause_text'].tolist()
test_clauses = splits['test']['clause_text'].tolist()
val_risk_labels = trainer.risk_discovery.get_risk_labels(val_clauses)
test_risk_labels = trainer.risk_discovery.get_risk_labels(test_clauses)
val_dataset = LegalClauseDataset(
clauses=val_clauses,
risk_labels=val_risk_labels,
severity_scores=trainer._generate_synthetic_scores(val_clauses, 'severity'),
importance_scores=trainer._generate_synthetic_scores(val_clauses, 'importance'),
tokenizer=trainer.tokenizer,
max_length=config.max_sequence_length
)
test_dataset = LegalClauseDataset(
clauses=test_clauses,
risk_labels=test_risk_labels,
severity_scores=trainer._generate_synthetic_scores(test_clauses, 'severity'),
importance_scores=trainer._generate_synthetic_scores(test_clauses, 'importance'),
tokenizer=trainer.tokenizer,
max_length=config.max_sequence_length
)
val_loader = DataLoader(val_dataset, batch_size=config.batch_size, shuffle=False, collate_fn=collate_batch)
test_loader = DataLoader(test_dataset, batch_size=config.batch_size, shuffle=False, collate_fn=collate_batch)
print(f"β
Data loaded: {len(val_dataset)} val, {len(test_dataset)} test samples")
# Initialize calibration framework
print("\n" + "=" * 80)
print("π‘οΈ PHASE 1: CALIBRATION")
print("=" * 80)
calibrator = CalibrationFramework(model, config.device)
# Calculate pre-calibration metrics
print("\nπ Pre-calibration metrics:")
ece_before = calibrator.calculate_ece(test_loader)
mce_before = calibrator.calculate_mce(test_loader)
# Apply temperature scaling
print("\nπ§ Calibrating model...")
optimal_temp = calibrator.temperature_scaling(val_loader)
# Calculate post-calibration metrics
print("\nπ Post-calibration metrics:")
ece_after = calibrator.calculate_ece(test_loader)
mce_after = calibrator.calculate_mce(test_loader)
# Save calibration results
print("\n" + "=" * 80)
print("πΎ SAVING RESULTS")
print("=" * 80)
calibration_results = {
'calibration_date': datetime.now().strftime('%Y-%m-%d %H:%M:%S'),
'optimal_temperature': optimal_temp,
'metrics': {
'pre_calibration': {
'ece': float(ece_before),
'mce': float(mce_before)
},
'post_calibration': {
'ece': float(ece_after),
'mce': float(mce_after)
},
'improvement': {
'ece': float(ece_before - ece_after),
'mce': float(mce_before - mce_after)
}
}
}
results_path = os.path.join(config.checkpoint_dir, 'calibration_results.json')
with open(results_path, 'w') as f:
json.dump(calibration_results, f, indent=2)
print(f"β
Results saved to: {results_path}")
# Save calibrated model
calibrated_model_path = os.path.join(config.model_save_path, 'calibrated_model.pt')
torch.save({
'model_state_dict': model.state_dict(),
'config': config,
'discovered_patterns': checkpoint['discovered_patterns'],
'temperature': optimal_temp,
'calibration_results': calibration_results
}, calibrated_model_path)
print(f"β
Calibrated model saved to: {calibrated_model_path}")
# Summary
print("\n" + "=" * 80)
print("β
CALIBRATION COMPLETE!")
print("=" * 80)
print(f"\nπ― Calibration Results:")
print(f" Optimal Temperature: {optimal_temp:.4f}")
print(f"\n ECE Improvement: {ece_before:.4f} β {ece_after:.4f} (Ξ {ece_before - ece_after:.4f})")
print(f" MCE Improvement: {mce_before:.4f} β {mce_after:.4f} (Ξ {mce_before - mce_after:.4f})")
if ece_after < 0.08:
print(f"\n β
Target ECE (<0.08) achieved!")
else:
print(f"\n β οΈ ECE slightly above target (0.08)")
print(f"\nπ― Next Steps:")
print(f" 1. Analyze calibration quality across risk categories")
print(f" 2. Compare with baseline methods")
print(f" 3. Generate final implementation report")
return calibrator, calibration_results
if __name__ == "__main__":
calibrator, results = main()
|