speech-emotion-recognition / src /genetic_algorithm.py
nguyennp86's picture
update GA model select
6146f6c
"""
Genetic Algorithm for feature selection and hyperparameter optimization
"""
import numpy as np
import random
import time
import warnings
from typing import Dict, List, Callable, Optional, Tuple
from joblib import Parallel, delayed
from xgboost import XGBClassifier
from lightgbm import LGBMClassifier
from sklearn.ensemble import GradientBoostingClassifier, AdaBoostClassifier
from sklearn.metrics import accuracy_score
import config
warnings.filterwarnings(
'ignore', message='X does not have valid feature names')
warnings.filterwarnings('ignore', category=UserWarning, module='sklearn')
class GeneticAlgorithm:
"""GA for optimizing features + hyperparameters + ensemble weights"""
def __init__(self, X: np.ndarray, y: np.ndarray,
n_features_to_select: int = 80,
skip_feature_selection: bool = False,
selected_models: List[str] = None):
"""
Initialize GA
Args:
X: Training data
y: Training labels
n_features_to_select: Number of features to select
skip_feature_selection: If True, use all features (only optimize hyperparams)
selected_models: List of models to train ['xgboost', 'lightgbm', 'gradientboosting', 'adaboost']
"""
self.X = X
self.y = y
self.n_features = X.shape[1]
self.skip_feature_selection = skip_feature_selection
# Model selection
if selected_models is None or len(selected_models) == 0:
self.selected_models = ['xgboost',
'lightgbm', 'gradientboosting', 'adaboost']
else:
self.selected_models = selected_models
self.n_models = len(self.selected_models)
if skip_feature_selection:
self.n_select = self.n_features
print(
f"✅ GA will optimize: HYPERPARAMETERS ONLY (using all {self.n_features} features)")
else:
if n_features_to_select > self.n_features:
print(
f"⚠️ Adjusted: {n_features_to_select}{self.n_features} features")
self.n_select = self.n_features
else:
self.n_select = n_features_to_select
print(
f"✅ GA will optimize: FEATURES ({self.n_select}/{self.n_features}) + HYPERPARAMETERS")
print(
f"✅ Training models: {', '.join(self.selected_models)} ({self.n_models} models)")
self.n_classes = len(np.unique(y))
# GA parameters from config
self.population_size = config.GA_CONFIG['population_size']
self.n_generations = config.GA_CONFIG['n_generations']
self.mutation_rate = config.GA_CONFIG['mutation_rate']
self.crossover_rate = config.GA_CONFIG['crossover_rate']
self.elite_size = config.GA_CONFIG['elite_size']
self.early_stopping_patience = config.GA_CONFIG['early_stopping_patience']
self.early_stopping_tolerance = config.GA_CONFIG['early_stopping_tolerance']
self.best_chromosome = None
self.best_fitness = 0
self.history = []
self.log_messages = []
def log(self, message: str):
"""Add log message with timestamp"""
timestamp = time.strftime("%H:%M:%S")
log_entry = f"[{timestamp}] {message}"
self.log_messages.append(log_entry)
print(log_entry)
def create_chromosome(self) -> Dict:
"""Create random chromosome"""
chromosome = {}
# Feature selection (skip if not optimizing features)
if self.skip_feature_selection:
chromosome['feature_indices'] = np.arange(self.n_features)
else:
n_to_select = min(self.n_select, self.n_features)
chromosome['feature_indices'] = np.sort(np.random.choice(
self.n_features, n_to_select, replace=False
))
# Add hyperparameters ONLY for selected models
for model_name in self.selected_models:
model_prefix = self._get_model_prefix(model_name)
if model_prefix in config.MODEL_HYPERPARAMS:
for param_name, param_values in config.MODEL_HYPERPARAMS[model_prefix].items():
key = f"{model_prefix}_{param_name}"
chromosome[key] = random.choice(param_values)
# Ensemble weights (for selected models only)
chromosome['weights'] = self._random_weights(self.n_models)
return chromosome
def _get_model_prefix(self, model_name: str) -> str:
"""Get model prefix for config lookup"""
prefix_map = {
'xgboost': 'xgb',
'lightgbm': 'lgbm',
'gradientboosting': 'gb',
'adaboost': 'ada'
}
return prefix_map.get(model_name, model_name)
def _random_weights(self, n: int) -> np.ndarray:
"""Generate n random weights that sum to 1"""
return np.random.dirichlet(np.ones(n))
def fitness(self, chromosome: Dict, X_train: np.ndarray, y_train: np.ndarray,
X_val: np.ndarray, y_val: np.ndarray) -> float:
"""Calculate fitness using validation accuracy"""
try:
feature_indices = chromosome['feature_indices']
X_train_selected = X_train[:, feature_indices]
X_val_selected = X_val[:, feature_indices]
models = []
# Train only selected models
for model_name in self.selected_models:
model = self._train_model(
model_name, chromosome,
X_train_selected, y_train
)
models.append(model)
# Ensemble prediction
predictions = [model.predict_proba(
X_val_selected) for model in models]
weights = chromosome['weights']
ensemble_proba = np.average(predictions, axis=0, weights=weights)
y_pred = np.argmax(ensemble_proba, axis=1)
accuracy = accuracy_score(y_val, y_pred)
return accuracy
except Exception as e:
print(f"⚠️ Error in fitness evaluation: {e}")
import traceback
traceback.print_exc()
return 0.0
def _train_model(self, model_name: str, chromosome: Dict, X_train: np.ndarray, y_train: np.ndarray):
"""Train a single model based on name and chromosome config"""
if model_name == 'xgboost':
model = XGBClassifier(
n_estimators=chromosome.get('xgb_n_estimators', 100),
max_depth=chromosome.get('xgb_max_depth', 6),
learning_rate=chromosome.get('xgb_learning_rate', 0.1),
subsample=chromosome.get('xgb_subsample', 0.8),
colsample_bytree=chromosome.get('xgb_colsample_bytree', 0.8),
min_child_weight=chromosome.get('xgb_min_child_weight', 1),
gamma=chromosome.get('xgb_gamma', 0),
objective='multi:softprob',
num_class=self.n_classes,
random_state=config.RANDOM_STATE,
n_jobs=-1,
verbosity=0
)
elif model_name == 'lightgbm':
model = LGBMClassifier(
n_estimators=chromosome.get('lgbm_n_estimators', 100),
num_leaves=chromosome.get('lgbm_num_leaves', 31),
learning_rate=chromosome.get('lgbm_learning_rate', 0.1),
min_child_samples=chromosome.get('lgbm_min_child_samples', 20),
subsample=chromosome.get('lgbm_subsample', 0.8),
colsample_bytree=chromosome.get('lgbm_colsample_bytree', 0.8),
reg_alpha=chromosome.get('lgbm_reg_alpha', 0),
reg_lambda=chromosome.get('lgbm_reg_lambda', 0),
objective='multiclass',
num_class=self.n_classes,
random_state=config.RANDOM_STATE,
n_jobs=-1,
verbose=-1,
force_col_wise=True
)
elif model_name == 'gradientboosting':
model = GradientBoostingClassifier(
n_estimators=chromosome.get('gb_n_estimators', 100),
max_depth=chromosome.get('gb_max_depth', 5),
learning_rate=chromosome.get('gb_learning_rate', 0.1),
subsample=chromosome.get('gb_subsample', 0.8),
min_samples_split=chromosome.get('gb_min_samples_split', 2),
min_samples_leaf=chromosome.get('gb_min_samples_leaf', 1),
random_state=config.RANDOM_STATE
)
elif model_name == 'adaboost':
model = AdaBoostClassifier(
n_estimators=chromosome.get('ada_n_estimators', 100),
learning_rate=chromosome.get('ada_learning_rate', 1.0),
algorithm=config.ADABOOST_ALGORITHM,
random_state=config.RANDOM_STATE
)
else:
raise ValueError(f"Unknown model: {model_name}")
model.fit(X_train, y_train)
return model
def crossover(self, parent1: Dict, parent2: Dict) -> Tuple[Dict, Dict]:
"""Crossover operation"""
if random.random() > self.crossover_rate:
return parent1.copy(), parent2.copy()
child1 = {}
child2 = {}
# Feature crossover
if self.skip_feature_selection:
child1['feature_indices'] = parent1['feature_indices'].copy()
child2['feature_indices'] = parent2['feature_indices'].copy()
else:
mask = np.random.rand(self.n_select) < 0.5
child1_features = np.where(
mask, parent1['feature_indices'], parent2['feature_indices'])
child2_features = np.where(
mask, parent2['feature_indices'], parent1['feature_indices'])
child1_features = np.unique(child1_features)
child2_features = np.unique(child2_features)
while len(child1_features) < self.n_select:
new_feat = random.randint(0, self.n_features - 1)
if new_feat not in child1_features:
child1_features = np.append(child1_features, new_feat)
while len(child2_features) < self.n_select:
new_feat = random.randint(0, self.n_features - 1)
if new_feat not in child2_features:
child2_features = np.append(child2_features, new_feat)
child1['feature_indices'] = np.sort(
child1_features[:self.n_select])
child2['feature_indices'] = np.sort(
child2_features[:self.n_select])
# Hyperparameter crossover
for key in parent1.keys():
if key != 'feature_indices':
if random.random() < 0.5:
child1[key] = parent1[key]
child2[key] = parent2[key]
else:
child1[key] = parent2[key]
child2[key] = parent1[key]
return child1, child2
def mutate(self, chromosome: Dict) -> Dict:
"""Mutation operation"""
mutated = chromosome.copy()
# Feature mutation
if not self.skip_feature_selection:
if random.random() < self.mutation_rate:
n_replace = random.randint(1, min(5, self.n_select))
indices_to_replace = np.random.choice(
self.n_select, n_replace, replace=False)
for idx in indices_to_replace:
new_feat = random.randint(0, self.n_features - 1)
while new_feat in mutated['feature_indices']:
new_feat = random.randint(0, self.n_features - 1)
mutated['feature_indices'][idx] = new_feat
mutated['feature_indices'] = np.sort(
mutated['feature_indices'])
# Hyperparameter mutation
if random.random() < self.mutation_rate:
param_keys = [k for k in chromosome.keys() if k not in [
'feature_indices', 'weights']]
if param_keys:
param_to_mutate = random.choice(param_keys)
temp = self.create_chromosome()
mutated[param_to_mutate] = temp[param_to_mutate]
# Weight mutation
if random.random() < self.mutation_rate:
mutated['weights'] = self._random_weights(self.n_models)
return mutated
def evaluate_population_parallel(self, population: List[Dict],
X_train: np.ndarray, y_train: np.ndarray,
X_val: np.ndarray, y_val: np.ndarray,
n_jobs: int = 2) -> List[float]:
"""Evaluate entire population in parallel"""
# Limit n_jobs to prevent resource exhaustion
safe_n_jobs = min(n_jobs, 4, len(population) // 2)
if safe_n_jobs < 1:
safe_n_jobs = 1
self.log(
f" Evaluating {len(population)} individuals (n_jobs={safe_n_jobs})...")
try:
fitness_scores = Parallel(
n_jobs=safe_n_jobs,
verbose=0,
backend='loky',
timeout=600
)(
delayed(self.fitness)(
chromosome, X_train, y_train, X_val, y_val)
for chromosome in population
)
except Exception as e:
self.log(f"⚠️ Parallel evaluation failed: {e}")
self.log(" Falling back to sequential evaluation...")
fitness_scores = []
for i, chromosome in enumerate(population):
if (i + 1) % 5 == 0:
self.log(
f" Progress: {i+1}/{len(population)} individuals")
score = self.fitness(chromosome, X_train,
y_train, X_val, y_val)
fitness_scores.append(score)
return fitness_scores
def evolve(self, X_train: np.ndarray, y_train: np.ndarray,
X_val: np.ndarray, y_val: np.ndarray,
progress_callback: Optional[Callable] = None,
n_jobs: int = 2) -> Dict:
"""Main GA evolution loop"""
self.log("="*70)
self.log("🧬 GENETIC ALGORITHM OPTIMIZATION")
self.log("="*70)
self.log(f"Population size: {self.population_size}")
self.log(f"Generations: {self.n_generations}")
self.log(
f"Feature selection: {'DISABLED (hyperparams only)' if self.skip_feature_selection else f'ENABLED ({self.n_select}/{self.n_features})'}")
self.log(
f"Selected models: {', '.join(self.selected_models)} ({self.n_models} models)")
self.log(f"Early stopping patience: {self.early_stopping_patience}")
self.log(f"Parallel jobs: {n_jobs}")
self.log("="*70)
population = [self.create_chromosome()
for _ in range(self.population_size)]
start_time = time.time()
no_improve_count = 0
for generation in range(self.n_generations):
try:
gen_start = time.time()
self.log(
f"\n📊 Generation {generation + 1}/{self.n_generations}")
# Parallel fitness evaluation
fitness_scores = self.evaluate_population_parallel(
population, X_train, y_train, X_val, y_val, n_jobs=n_jobs
)
# Validation check
if len(fitness_scores) != len(population):
self.log(
f"⚠️ Warning: Got {len(fitness_scores)} scores for {len(population)} individuals")
while len(fitness_scores) < len(population):
fitness_scores.append(0.0)
max_fitness = max(fitness_scores)
avg_fitness = np.mean(fitness_scores)
std_fitness = np.std(fitness_scores)
max_idx = fitness_scores.index(max_fitness)
# Track improvement
improved = False
if max_fitness > self.best_fitness + self.early_stopping_tolerance:
prev_best = self.best_fitness
self.best_fitness = max_fitness
self.best_chromosome = population[max_idx].copy()
no_improve_count = 0
improved = True
self.log(
f" ✨ NEW BEST: {max_fitness:.4f} (+{max_fitness - prev_best:.4f})")
else:
no_improve_count += 1
self.log(
f" → Best: {max_fitness:.4f} (no improvement, count={no_improve_count})")
# Log statistics
self.log(
f" Average: {avg_fitness:.4f} (σ={std_fitness:.4f})")
self.log(
f" Range: [{min(fitness_scores):.4f}, {max(fitness_scores):.4f}]")
gen_time = time.time() - gen_start
elapsed = time.time() - start_time
avg_gen_time = elapsed / (generation + 1)
eta = avg_gen_time * (self.n_generations - generation - 1)
self.log(
f" Time: {gen_time:.1f}s | Elapsed: {elapsed/60:.1f}min | ETA: {eta/60:.1f}min")
self.history.append({
'generation': generation + 1,
'best_fitness': max_fitness,
'avg_fitness': avg_fitness,
'std_fitness': std_fitness,
'time': gen_time,
'improved': improved
})
# Update progress callback
if progress_callback:
try:
progress_callback(
(generation + 1) / self.n_generations,
desc=f"Gen {generation+1}/{self.n_generations} | Best: {max_fitness:.4f} | Avg: {avg_fitness:.4f} | ETA: {eta/60:.0f}min"
)
except Exception as e:
self.log(f"⚠️ Progress callback failed: {e}")
# Early stopping check
if no_improve_count >= self.early_stopping_patience:
self.log(
f"\n🛑 EARLY STOPPING at generation {generation + 1}")
self.log(
f" No improvement for {self.early_stopping_patience} consecutive generations")
self.log(f" Best fitness: {self.best_fitness:.4f}")
break
# Explicit flush
import sys
sys.stdout.flush()
# Selection
self.log(f" Creating next generation...")
selected = []
for _ in range(self.population_size - self.elite_size):
tournament = random.sample(
list(zip(population, fitness_scores)), 3)
winner = max(tournament, key=lambda x: x[1])[0]
selected.append(winner)
elite_indices = np.argsort(fitness_scores)[-self.elite_size:]
elite = [population[i] for i in elite_indices]
# Crossover & Mutation
offspring = []
for i in range(0, len(selected), 2):
if i + 1 < len(selected):
child1, child2 = self.crossover(
selected[i], selected[i+1])
offspring.append(self.mutate(child1))
offspring.append(self.mutate(child2))
population = elite + \
offspring[:self.population_size - self.elite_size]
self.log(f" ✓ Generation {generation + 1} complete")
except KeyboardInterrupt:
self.log("\n⚠️ Training interrupted by user")
break
except Exception as e:
self.log(f"\n❌ Error in generation {generation + 1}: {e}")
import traceback
self.log(traceback.format_exc())
if generation == 0:
self.log("❌ First generation failed, aborting")
return None
else:
self.log("⚠️ Attempting to continue...")
continue
total_time = time.time() - start_time
self.log("\n" + "="*70)
self.log("✅ GA OPTIMIZATION COMPLETE")
self.log("="*70)
self.log(f"Final best fitness: {self.best_fitness:.4f}")
self.log(
f"Total generations: {len(self.history)}/{self.n_generations}")
self.log(f"Total time: {total_time/60:.1f} minutes")
if len(self.history) > 0:
self.log(
f"Average time per generation: {total_time/len(self.history):.1f}s")
self.log("="*70)
if self.best_chromosome is None:
self.log(
"⚠️ Warning: No improvement found, using best from final generation")
fitness_scores = self.evaluate_population_parallel(
population, X_train, y_train, X_val, y_val, n_jobs=n_jobs
)
max_idx = fitness_scores.index(max(fitness_scores))
self.best_chromosome = population[max_idx].copy()
self.best_fitness = fitness_scores[max_idx]
return self.best_chromosome