Hyperparameter Optimization

import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import itertools
 
def build_and_evaluate(units, lr, X_train, y_train, X_val, y_val, epochs=15):
    model = keras.Sequential([
        layers.Input(shape=(X_train.shape[1], X_train.shape[2])),
        layers.LSTM(units),
        layers.Dense(1),
    ])
    model.compile(optimizer=keras.optimizers.Adam(lr), loss="mse")
    model.fit(X_train, y_train, epochs=epochs, verbose=0,
              callbacks=[keras.callbacks.EarlyStopping(patience=3, restore_best_weights=True)])
    val_loss = model.evaluate(X_val, y_val, verbose=0)
    return val_loss
 
units_grid = [16, 32, 64]
lr_grid    = [1e-3, 5e-4, 1e-4]
results = []
 
# itertools.product দিয়ে সব কম্বিনেশন তৈরি করা
for units, lr in itertools.product(units_grid, lr_grid):
    loss = build_and_evaluate(units, lr, X_train, y_train, X_val, y_val)
    results.append((units, lr, loss))
    print(f"units={units}, lr={lr:.0e} → val_loss={loss:.4f}")
 
best = min(results, key=lambda x: x[2])
print(f"\nBest: units={best[0]}, lr={best[1]:.0e}, val_loss={best[2]:.4f}")

import numpy as np
 
n_trials = 20
best_loss = float("inf")
best_config = None
 
for trial in range(n_trials):
    units = int(np.random.choice([16, 32, 64, 128]))
    lr    = float(10 ** np.random.uniform(-4, -2))   # log-uniform: 0.01 থেকে 0.0001
    loss  = build_and_evaluate(units, lr, X_train, y_train, X_val, y_val)
    print(f"trial {trial:2d}: units={units}, lr={lr:.2e} → val_loss={loss:.4f}")
    if loss < best_loss:
        best_loss, best_config = loss, (units, lr)
 
print(f"\nBest: units={best_config[0]}, lr={best_config[1]:.2e}, val_loss={best_loss:.4f}")

import optuna
 
def objective(trial):
    units = trial.suggest_categorical("units", [16, 32, 64, 128])
    lr    = trial.suggest_float("lr", 1e-4, 1e-2, log=True)
    dropout = trial.suggest_float("dropout", 0.0, 0.5)
 
    model = keras.Sequential([
        layers.Input(shape=(X_train.shape[1], X_train.shape[2])),
        layers.LSTM(units, dropout=dropout),
        layers.Dense(1),
    ])
    model.compile(optimizer=keras.optimizers.Adam(lr), loss="mse")
    model.fit(X_train, y_train, epochs=20, verbose=0,
              callbacks=[keras.callbacks.EarlyStopping(patience=3)])
    return model.evaluate(X_val, y_val, verbose=0)
 
study = optuna.create_study(direction="minimize")
study.optimize(objective, n_trials=30, show_progress_bar=True)
print("Best params:", study.best_params)