File size: 10,805 Bytes
94cb318 | 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 | """Automated FL Hyperparameter Optimizer (Optuna / Bayesian TPE).
Executes Bayesian optimization over federated learning hyperparameters (learning rate, local epochs,
DP clip norm, noise multiplier, staleness decay, FedProx penalty mu) tailored to Non-IID bank data distributions.
"""
from __future__ import annotations
import logging
import time
from typing import Any
import numpy as np
import optuna
from app.application.services.fl_dirichlet_partitioner import DirichletPartitioner
from app.application.services.fl_engine import FederatedLearningEngine as FLEngine
from app.domain.value_objects import ModelWeights
logger = logging.getLogger(__name__)
# Suppress noisy Optuna verbosity by default
optuna.logging.set_verbosity(optuna.logging.WARNING)
class FLHyperparameterOptimizer:
"""Automated Optuna Bayesian TPE Hyperparameter Tuning Engine for FL."""
def __init__(
self,
study_name: str = "fl_hpo_study",
dirichlet_alpha: float = 0.5,
num_clients: int = 3,
num_rounds: int = 5,
seed: int = 42,
) -> None:
self.study_name = study_name
self.dirichlet_alpha = dirichlet_alpha
self.num_clients = num_clients
self.num_rounds = num_rounds
self.seed = seed
# Create Optuna TPE sampler and MedianPruner
self.sampler = optuna.samplers.TPESampler(seed=seed)
self.pruner = optuna.pruners.MedianPruner(n_startup_trials=2, n_warmup_steps=1)
self.study = optuna.create_study(
study_name=study_name,
direction="maximize",
sampler=self.sampler,
pruner=self.pruner,
)
def _generate_synthetic_bank_data(self) -> tuple[np.ndarray, np.ndarray]:
"""Generate reproducible synthetic fraud transaction features and labels."""
rng = np.random.default_rng(self.seed)
n_samples = 600
n_features = 10
# Class 0: Legitimate, Class 1: Fraud
X0 = rng.normal(loc=0.0, scale=1.0, size=(int(n_samples * 0.9), n_features))
y0 = np.zeros(int(n_samples * 0.9))
X1 = rng.normal(loc=2.5, scale=1.2, size=(int(n_samples * 0.1), n_features))
y1 = np.ones(int(n_samples * 0.1))
X = np.vstack([X0, X1])
y = np.concatenate([y0, y1])
perm = rng.permutation(len(y))
return X[perm], y[perm]
def objective(self, trial: optuna.Trial) -> float:
"""Evaluate a single trial set of candidate FL hyperparameters using real PyTorch training."""
# 1. Suggest hyperparameters
lr = trial.suggest_float("learning_rate", 1e-4, 1e-1, log=True)
local_epochs = trial.suggest_int("local_epochs", 1, 5)
dp_clip_norm = trial.suggest_float("dp_clip_norm", 0.1, 5.0)
dp_noise_multiplier = trial.suggest_float("dp_noise_multiplier", 0.1, 2.0)
staleness_gamma = trial.suggest_float("staleness_gamma", 0.1, 3.0)
fedprox_mu = trial.suggest_float("fedprox_mu", 1e-3, 1.0, log=True)
batch_size = trial.suggest_categorical("batch_size", [16, 32, 64])
# 2. Generate dataset and partition across bank nodes using Dirichlet Dir(alpha)
X, y = self._generate_synthetic_bank_data()
split_idx = int(len(X) * 0.8)
X_train, X_val = X[:split_idx], X[split_idx:]
y_train, y_val = y[:split_idx], y[split_idx:]
client_datasets = DirichletPartitioner.partition_dataset(
features=X_train,
labels=y_train,
num_clients=self.num_clients,
alpha=self.dirichlet_alpha,
seed=self.seed + trial.number,
)
from app.application.services.model_service import ModelService
from app.application.services.privacy_service import PrivacyService
from app.config import get_settings
settings = get_settings()
model_service = ModelService(settings)
privacy_service = PrivacyService()
fl_engine = FLEngine(settings, model_service, privacy_service)
# Initialize global model with reproducible seed
rng = np.random.default_rng(self.seed)
global_model = model_service.create_model(dp_compatible=True)
global_weights = model_service.get_parameters(global_model)
final_score = 0.5
# 3. Simulate federated rounds with genuine local PyTorch training
for round_idx in range(1, self.num_rounds + 1):
client_updates = []
client_samples = []
# Staleness attenuation factor S(tau) = (1 + 0.1 * round)^(-gamma)
staleness_weight = (1.0 + (round_idx - 1) * 0.1) ** (-staleness_gamma)
for client_idx, (X_i, y_i) in enumerate(client_datasets):
# Instantiate client model matching global weights
local_model = model_service.create_model(dp_compatible=True)
local_model = model_service.set_parameters(local_model, global_weights)
# Train genuine local PyTorch model with FedProx proximal regularization
local_model, _, _ = model_service.train_local(
model=local_model,
X_train=X_i,
y_train=y_i,
epochs=local_epochs,
batch_size=batch_size,
learning_rate=lr,
global_weights=global_weights,
fedprox_mu=fedprox_mu,
)
updated_weights = model_service.get_parameters(local_model)
# Calibrate Differential Privacy noise perturbation on local weights
dp_noise = rng.normal(
0.0,
(dp_noise_multiplier * dp_clip_norm) / np.sqrt(max(len(X_i), 1)),
len(updated_weights.flat_weights),
)
noisy_flat = (np.array(updated_weights.flat_weights) + dp_noise).tolist()
noisy_weights = ModelWeights(
layer_shapes=updated_weights.layer_shapes,
flat_weights=noisy_flat,
)
client_updates.append(noisy_weights)
# Attenuate sample influence by staleness factor
effective_samples = max(1, int(len(X_i) * staleness_weight))
client_samples.append(effective_samples)
# Aggregate round updates via sample-weighted consensus
global_weights = fl_engine.aggregate_parameters(client_updates, client_samples)
global_model = model_service.set_parameters(global_model, global_weights)
# Evaluate round performance on real holdout validation dataset
eval_metrics = model_service.evaluate(global_model, X_val, y_val)
raw_round_auc = eval_metrics.get("auc_roc")
if raw_round_auc is None:
logger.warning(
"Trial %d produced undefined validation AUC at round %d; pruning trial explicitly",
trial.number,
round_idx,
)
raise optuna.TrialPruned(f"Undefined validation AUC at round {round_idx}")
round_auc = float(raw_round_auc)
final_score = round_auc
# Report step metric to Optuna for early MedianPruner decisions
trial.report(round_auc, round_idx)
if trial.should_prune():
logger.info(
"Trial %d pruned at round %d (validation AUC: %.4f)",
trial.number,
round_idx,
round_auc,
)
raise optuna.TrialPruned()
return final_score
def run_optimization(self, n_trials: int = 5, timeout: float | None = 60.0) -> dict[str, Any]:
"""Run Optuna optimization study over n_trials with graceful pruning handling."""
start_time = time.perf_counter()
logger.info(
"Starting Optuna FL hyperparameter optimization study '%s' (Dirichlet alpha=%.2f, trials=%d)...",
self.study_name,
self.dirichlet_alpha,
n_trials,
)
self.study.optimize(self.objective, n_trials=n_trials, timeout=timeout)
duration = (time.perf_counter() - start_time) * 1000
default_params = {
"learning_rate": 0.01,
"local_epochs": 2,
"dp_clip_norm": 1.0,
"dp_noise_multiplier": 0.5,
"staleness_gamma": 0.5,
"fedprox_mu": 0.01,
"batch_size": 32,
}
# Safe best trial extraction (handles all-pruned or empty study cases)
completed_trials = [
t for t in self.study.trials if t.state == optuna.trial.TrialState.COMPLETE
]
if completed_trials:
best_t = self.study.best_trial
best_trial_number = best_t.number
best_value = float(best_t.value) if best_t.value is not None else 0.5
best_params = best_t.params if best_t.params else default_params
elif self.study.trials:
best_t = max(
self.study.trials,
key=lambda t: max(t.intermediate_values.values()) if t.intermediate_values else -1.0,
)
best_trial_number = best_t.number
best_value = (
max(best_t.intermediate_values.values()) if best_t.intermediate_values else 0.5
)
best_params = best_t.params if best_t.params else default_params
else:
best_trial_number = -1
best_value = 0.5
best_params = default_params
try:
if len(completed_trials) >= 2:
importances = optuna.importance.get_param_importances(self.study)
else:
importances = {}
except Exception as exc:
logger.debug("Could not compute Optuna parameter importances: %s", exc)
importances = {}
results = {
"study_name": self.study_name,
"dirichlet_alpha": self.dirichlet_alpha,
"best_trial_number": best_trial_number,
"best_value": round(best_value, 4),
"best_params": best_params,
"param_importances": importances,
"total_trials": len(self.study.trials),
"completed_trials": len(completed_trials),
"pruned_trials": len(
[t for t in self.study.trials if t.state == optuna.trial.TrialState.PRUNED]
),
"duration_ms": round(duration, 2),
}
logger.info(
"Completed Optuna study '%s' in %.2fms. Best trial #%d (AUC: %.4f)",
self.study_name,
duration,
best_trial_number,
best_value,
)
return results
|