import itertools from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, Dropout from tensorflow.keras.datasets import mnist from tensorflow.keras.utils import to_categorical # データ準備 (X_train, y_train), (X_test, y_test) = mnist.load_data() X_train = X_train.reshape(-1, 784).astype(‘float32’) / 255.0 X_test = X_test.reshape(-1, 784).astype(‘float32’) / 255.0 y_train = to_categorical(y_train, 10) y_test = to_categorical(y_test, 10) # ハイパーパラメータの候補 param_grid = { ‘units’: [64, 128], ‘dropout’: [0.2, 0.3], ‘learning_rate’: [0.001, 0.01] } # 全組み合わせを生成 keys = param_grid.keys() combinations = list(itertools.product(*param_grid.values())) best_acc = 0 best_params = None for combo in combinations: params = dict(zip(keys, combo)) print(f”Testing: {params}”) # モデル構築 model = Sequential([ Dense(params[‘units’], activation=’relu’, input_shape=(784,)), Dropout(params[‘dropout’]), Dense(params[‘units’], activation=’relu’), Dropout(params[‘dropout’]), Dense(10, activation=’softmax’) ]) from tensorflow.keras.optimizers import Adam model.compile( optimizer=Adam(learning_rate=params[‘learning_rate’]), loss=’categorical_crossentropy’, metrics=[‘accuracy’] ) # 学習（少ないエポックで評価） model.fit(X_train, y_train, epochs=5, batch_size=64, validation_split=0.1, verbose=0) # 評価 _, acc = model.evaluate(X_test, y_test, verbose=0) print(f” Accuracy: {acc:.4f}”) if acc > best_acc: best_acc = acc best_params = params print(f”\nBest: {best_params}, Accuracy: {best_acc:.4f}”)

import numpy as np from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, Dropout from tensorflow.keras.optimizers import Adam # ハイパーパラメータの範囲 param_distributions = { ‘units’: [32, 64, 128, 256], ‘dropout’: (0.1, 0.5), # 連続値 ‘learning_rate’: (1e-4, 1e-2), # 連続値（対数スケール） ‘batch_size’: [32, 64, 128] } n_trials = 10 best_acc = 0 best_params = None for trial in range(n_trials): # ランダムにパラメータを選択 params = { ‘units’: np.random.choice(param_distributions[‘units’]), ‘dropout’: np.random.uniform(*param_distributions[‘dropout’]), ‘learning_rate’: 10 ** np.random.uniform( np.log10(param_distributions[‘learning_rate’][0]), np.log10(param_distributions[‘learning_rate’][1]) ), ‘batch_size’: np.random.choice(param_distributions[‘batch_size’]) } print(f”Trial {trial+1}/{n_trials}: {params}”) # モデル構築・学習・評価 model = Sequential([ Dense(params[‘units’], activation=’relu’, input_shape=(784,)), Dropout(params[‘dropout’]), Dense(params[‘units’], activation=’relu’), Dropout(params[‘dropout’]), Dense(10, activation=’softmax’) ]) model.compile( optimizer=Adam(learning_rate=params[‘learning_rate’]), loss=’categorical_crossentropy’, metrics=[‘accuracy’] ) model.fit(X_train, y_train, epochs=5, batch_size=int(params[‘batch_size’]), validation_split=0.1, verbose=0) _, acc = model.evaluate(X_test, y_test, verbose=0) print(f” Accuracy: {acc:.4f}”) if acc > best_acc: best_acc = acc best_params = params print(f”\nBest: {best_params}”) print(f”Best Accuracy: {best_acc:.4f}”)

# インストール（Google Colabで実行） !pip install keras-tuner -q

import keras_tuner as kt from tensorflow import keras from tensorflow.keras import layers # ===== モデル構築関数 ===== def build_model(hp): “”” hp: HyperParametersオブジェクト ハイパーパラメータの範囲を指定 “”” model = keras.Sequential() # ユニット数を探索（32〜256、32刻み） units = hp.Int(‘units’, min_value=32, max_value=256, step=32) model.add(layers.Dense(units, activation=’relu’, input_shape=(784,))) # ドロップアウト率を探索（0.1〜0.5） dropout = hp.Float(‘dropout’, min_value=0.1, max_value=0.5, step=0.1) model.add(layers.Dropout(dropout)) # 2層目 model.add(layers.Dense(units, activation=’relu’)) model.add(layers.Dropout(dropout)) # 出力層 model.add(layers.Dense(10, activation=’softmax’)) # 学習率を探索（対数スケール） learning_rate = hp.Float(‘learning_rate’, min_value=1e-4, max_value=1e-2, sampling=’log’) model.compile( optimizer=keras.optimizers.Adam(learning_rate=learning_rate), loss=’categorical_crossentropy’, metrics=[‘accuracy’] ) return model

# ===== Random Searchチューナー ===== tuner = kt.RandomSearch( build_model, objective=’val_accuracy’, # 最大化する指標 max_trials=20, # 最大試行回数 executions_per_trial=1, # 各試行の実行回数 directory=’tuner_results’, # 結果の保存先 project_name=’mnist_tuning’ ) # 探索空間の確認 tuner.search_space_summary() # ===== 探索実行 ===== tuner.search( X_train, y_train, epochs=10, validation_split=0.2, callbacks=[keras.callbacks.EarlyStopping(patience=3)] ) # ===== 結果の確認 ===== # 最良のハイパーパラメータ best_hps = tuner.get_best_hyperparameters(num_trials=1)[0] print(f””” 最良のハイパーパラメータ: – units: {best_hps.get(‘units’)} – dropout: {best_hps.get(‘dropout’):.2f} – learning_rate: {best_hps.get(‘learning_rate’):.6f} “””) # 最良のモデルを取得 best_model = tuner.get_best_models(num_models=1)[0] test_loss, test_acc = best_model.evaluate(X_test, y_test) print(f”テスト精度: {test_acc:.4f}”)

Search space summary:
- units: Int(32, 256, step=32)
- dropout: Float(0.1, 0.5, step=0.1)
- learning_rate: Float(1e-4, 1e-2, sampling=log)

Trial 20/20 complete
Best val_accuracy: 0.9823

最良のハイパーパラメータ:
- units: 128
- dropout: 0.20
- learning_rate: 0.001234

テスト精度: 0.9801

# 学習率の探索範囲（対数スケール） # 1e-5, 1e-4, 1e-3, 1e-2, 1e-1 を試す learning_rates = [1e-5, 1e-4, 1e-3, 1e-2, 1e-1] for lr in learning_rates: model = create_model() model.compile(optimizer=Adam(learning_rate=lr), …) history = model.fit(X_train, y_train, epochs=5, …) # 学習曲線を確認 # 発散する → 学習率が高すぎる # 学習が遅すぎる → 学習率が低すぎる