Introdução ao sintonizador Keras

Veja no TensorFlow.org Executar no Google Colab Ver fonte no GitHub Baixar caderno

Visão geral

O Keras Tuner é uma biblioteca que ajuda você a escolher o conjunto ideal de hiperparâmetros para seu programa TensorFlow. O processo de selecionar o conjunto certo de hiperparâmetros para seu aplicativo de aprendizado de máquina (ML) é chamado de ajuste de hiperparâmetro ou hypertuning .

Os hiperparâmetros são as variáveis ​​que governam o processo de treinamento e a topologia de um modelo de ML. Essas variáveis ​​permanecem constantes ao longo do processo de treinamento e impactam diretamente no desempenho do seu programa de ML. Os hiperparâmetros são de dois tipos:

  1. Hiperparâmetros do modelo que influenciam a seleção do modelo, como o número e a largura das camadas ocultas
  2. Hiperparâmetros do algoritmo que influenciam a velocidade e a qualidade do algoritmo de aprendizado, como a taxa de aprendizado para Stochastic Gradient Descent (SGD) e o número de vizinhos mais próximos para o classificador ak Nearest Neighbors (KNN)

Neste tutorial, você usará o Keras Tuner para realizar o hypertuning para um aplicativo de classificação de imagens.

Configurar

import tensorflow as tf
from tensorflow import keras

Instale e importe o Keras Tuner.

pip install -q -U keras-tuner
import keras_tuner as kt

Baixe e prepare o conjunto de dados

Neste tutorial, você usará o Keras Tuner para encontrar os melhores hiperparâmetros para um modelo de aprendizado de máquina que classifica imagens de roupas do conjunto de dados Fashion MNIST .

Carregue os dados.

(img_train, label_train), (img_test, label_test) = keras.datasets.fashion_mnist.load_data()
# Normalize pixel values between 0 and 1
img_train = img_train.astype('float32') / 255.0
img_test = img_test.astype('float32') / 255.0

Defina o modelo

Ao construir um modelo para hypertuning, você também define o espaço de pesquisa do hiperparâmetro além da arquitetura do modelo. O modelo que você configurou para hypertuning é chamado de hipermodelo .

Você pode definir um hipermodelo por meio de duas abordagens:

  • Usando uma função de construtor de modelo
  • Subclassificando a classe HyperModel da API Keras Tuner

Você também pode usar duas classes HyperModel predefinidas - HyperXception e HyperResNet para aplicativos de visão computacional.

Neste tutorial, você usa uma função de construtor de modelo para definir o modelo de classificação de imagem. A função de construtor de modelo retorna um modelo compilado e usa hiperparâmetros que você define sequencialmente para ajustar o modelo.

def model_builder(hp):
  model = keras.Sequential()
  model.add(keras.layers.Flatten(input_shape=(28, 28)))

  # Tune the number of units in the first Dense layer
  # Choose an optimal value between 32-512
  hp_units = hp.Int('units', min_value=32, max_value=512, step=32)
  model.add(keras.layers.Dense(units=hp_units, activation='relu'))
  model.add(keras.layers.Dense(10))

  # Tune the learning rate for the optimizer
  # Choose an optimal value from 0.01, 0.001, or 0.0001
  hp_learning_rate = hp.Choice('learning_rate', values=[1e-2, 1e-3, 1e-4])

  model.compile(optimizer=keras.optimizers.Adam(learning_rate=hp_learning_rate),
                loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                metrics=['accuracy'])

  return model

Instancie o sintonizador e execute o hypertuning

Instancie o sintonizador para realizar o hypertuning. O Keras Tuner tem quatro sintonizadores disponíveis - RandomSearch , Hyperband , BayesianOptimization e Sklearn . Neste tutorial, você usa o sintonizador Hyperband .

Para instanciar o sintonizador Hyperband, você deve especificar o hipermodelo, o objective a ser otimizado e o número máximo de épocas para treinar ( max_epochs ).

tuner = kt.Hyperband(model_builder,
                     objective='val_accuracy',
                     max_epochs=10,
                     factor=3,
                     directory='my_dir',
                     project_name='intro_to_kt')

O algoritmo de ajuste Hyperband usa alocação adaptável de recursos e parada antecipada para convergir rapidamente em um modelo de alto desempenho. Isso é feito usando um suporte de estilo de campeonato esportivo. O algoritmo treina um grande número de modelos por algumas épocas e leva apenas a metade dos modelos com melhor desempenho para a próxima rodada. A hiperbanda determina o número de modelos a serem treinados em um colchete calculando 1 + factor logarítmico ( max_epochs ) e arredondando-o para o inteiro mais próximo.

Crie um retorno de chamada para interromper o treinamento antes de atingir um determinado valor para a perda de validação.

stop_early = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=5)

Execute a pesquisa de hiperparâmetros. Os argumentos para o método de pesquisa são os mesmos usados ​​para tf.keras.model.fit , além do retorno de chamada acima.

tuner.search(img_train, label_train, epochs=50, validation_split=0.2, callbacks=[stop_early])

# Get the optimal hyperparameters
best_hps=tuner.get_best_hyperparameters(num_trials=1)[0]

print(f"""
The hyperparameter search is complete. The optimal number of units in the first densely-connected
layer is {best_hps.get('units')} and the optimal learning rate for the optimizer
is {best_hps.get('learning_rate')}.
""")
Trial 30 Complete [00h 00m 35s]
val_accuracy: 0.8925833106040955

Best val_accuracy So Far: 0.8925833106040955
Total elapsed time: 00h 07m 26s
INFO:tensorflow:Oracle triggered exit

The hyperparameter search is complete. The optimal number of units in the first densely-connected
layer is 320 and the optimal learning rate for the optimizer
is 0.001.

Treine o modelo

Encontre o número ótimo de épocas para treinar o modelo com os hiperparâmetros obtidos na busca.

# Build the model with the optimal hyperparameters and train it on the data for 50 epochs
model = tuner.hypermodel.build(best_hps)
history = model.fit(img_train, label_train, epochs=50, validation_split=0.2)

val_acc_per_epoch = history.history['val_accuracy']
best_epoch = val_acc_per_epoch.index(max(val_acc_per_epoch)) + 1
print('Best epoch: %d' % (best_epoch,))
Epoch 1/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.4988 - accuracy: 0.8232 - val_loss: 0.4142 - val_accuracy: 0.8517
Epoch 2/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.3717 - accuracy: 0.8646 - val_loss: 0.3437 - val_accuracy: 0.8773
Epoch 3/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.3317 - accuracy: 0.8779 - val_loss: 0.3806 - val_accuracy: 0.8639
Epoch 4/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.3079 - accuracy: 0.8867 - val_loss: 0.3321 - val_accuracy: 0.8801
Epoch 5/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2882 - accuracy: 0.8943 - val_loss: 0.3313 - val_accuracy: 0.8806
Epoch 6/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2727 - accuracy: 0.8977 - val_loss: 0.3152 - val_accuracy: 0.8857
Epoch 7/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2610 - accuracy: 0.9016 - val_loss: 0.3225 - val_accuracy: 0.8873
Epoch 8/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2474 - accuracy: 0.9060 - val_loss: 0.3198 - val_accuracy: 0.8867
Epoch 9/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2385 - accuracy: 0.9105 - val_loss: 0.3266 - val_accuracy: 0.8822
Epoch 10/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2295 - accuracy: 0.9142 - val_loss: 0.3382 - val_accuracy: 0.8835
Epoch 11/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2170 - accuracy: 0.9185 - val_loss: 0.3215 - val_accuracy: 0.8885
Epoch 12/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2102 - accuracy: 0.9202 - val_loss: 0.3194 - val_accuracy: 0.8923
Epoch 13/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2036 - accuracy: 0.9235 - val_loss: 0.3176 - val_accuracy: 0.8901
Epoch 14/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1955 - accuracy: 0.9272 - val_loss: 0.3269 - val_accuracy: 0.8912
Epoch 15/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1881 - accuracy: 0.9292 - val_loss: 0.3391 - val_accuracy: 0.8878
Epoch 16/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1821 - accuracy: 0.9321 - val_loss: 0.3272 - val_accuracy: 0.8920
Epoch 17/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1771 - accuracy: 0.9332 - val_loss: 0.3536 - val_accuracy: 0.8876
Epoch 18/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1697 - accuracy: 0.9363 - val_loss: 0.3395 - val_accuracy: 0.8927
Epoch 19/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1652 - accuracy: 0.9374 - val_loss: 0.3464 - val_accuracy: 0.8937
Epoch 20/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1606 - accuracy: 0.9392 - val_loss: 0.3576 - val_accuracy: 0.8888
Epoch 21/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1539 - accuracy: 0.9417 - val_loss: 0.3724 - val_accuracy: 0.8867
Epoch 22/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1503 - accuracy: 0.9435 - val_loss: 0.3607 - val_accuracy: 0.8954
Epoch 23/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1450 - accuracy: 0.9454 - val_loss: 0.3525 - val_accuracy: 0.8919
Epoch 24/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1398 - accuracy: 0.9473 - val_loss: 0.3745 - val_accuracy: 0.8919
Epoch 25/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1370 - accuracy: 0.9478 - val_loss: 0.3616 - val_accuracy: 0.8941
Epoch 26/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1334 - accuracy: 0.9498 - val_loss: 0.3866 - val_accuracy: 0.8956
Epoch 27/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1282 - accuracy: 0.9519 - val_loss: 0.3947 - val_accuracy: 0.8924
Epoch 28/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1254 - accuracy: 0.9538 - val_loss: 0.4223 - val_accuracy: 0.8870
Epoch 29/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1222 - accuracy: 0.9536 - val_loss: 0.3805 - val_accuracy: 0.8898
Epoch 30/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1179 - accuracy: 0.9546 - val_loss: 0.4052 - val_accuracy: 0.8942
Epoch 31/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1162 - accuracy: 0.9560 - val_loss: 0.3909 - val_accuracy: 0.8955
Epoch 32/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.1152 - accuracy: 0.9572 - val_loss: 0.4160 - val_accuracy: 0.8908
Epoch 33/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1100 - accuracy: 0.9583 - val_loss: 0.4280 - val_accuracy: 0.8938
Epoch 34/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1055 - accuracy: 0.9603 - val_loss: 0.4148 - val_accuracy: 0.8963
Epoch 35/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1044 - accuracy: 0.9606 - val_loss: 0.4302 - val_accuracy: 0.8921
Epoch 36/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1046 - accuracy: 0.9605 - val_loss: 0.4205 - val_accuracy: 0.8947
Epoch 37/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0993 - accuracy: 0.9621 - val_loss: 0.4551 - val_accuracy: 0.8875
Epoch 38/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0972 - accuracy: 0.9635 - val_loss: 0.4622 - val_accuracy: 0.8914
Epoch 39/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0951 - accuracy: 0.9642 - val_loss: 0.4423 - val_accuracy: 0.8950
Epoch 40/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0947 - accuracy: 0.9637 - val_loss: 0.4498 - val_accuracy: 0.8948
Epoch 41/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0876 - accuracy: 0.9675 - val_loss: 0.4694 - val_accuracy: 0.8959
Epoch 42/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0902 - accuracy: 0.9657 - val_loss: 0.4778 - val_accuracy: 0.8938
Epoch 43/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0876 - accuracy: 0.9676 - val_loss: 0.4716 - val_accuracy: 0.8911
Epoch 44/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0884 - accuracy: 0.9674 - val_loss: 0.4827 - val_accuracy: 0.8918
Epoch 45/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0764 - accuracy: 0.9715 - val_loss: 0.5008 - val_accuracy: 0.8953
Epoch 46/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0823 - accuracy: 0.9695 - val_loss: 0.5157 - val_accuracy: 0.8874
Epoch 47/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0789 - accuracy: 0.9704 - val_loss: 0.5198 - val_accuracy: 0.8910
Epoch 48/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0778 - accuracy: 0.9716 - val_loss: 0.5031 - val_accuracy: 0.8932
Epoch 49/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0747 - accuracy: 0.9718 - val_loss: 0.4982 - val_accuracy: 0.8953
Epoch 50/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0786 - accuracy: 0.9706 - val_loss: 0.5198 - val_accuracy: 0.8976
Best epoch: 50

Reinstanciar o hipermodelo e treiná-lo com o número ideal de épocas de cima.

hypermodel = tuner.hypermodel.build(best_hps)

# Retrain the model
hypermodel.fit(img_train, label_train, epochs=best_epoch, validation_split=0.2)
Epoch 1/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.4987 - accuracy: 0.8236 - val_loss: 0.4065 - val_accuracy: 0.8488
Epoch 2/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.3738 - accuracy: 0.8652 - val_loss: 0.3847 - val_accuracy: 0.8613
Epoch 3/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.3344 - accuracy: 0.8775 - val_loss: 0.3568 - val_accuracy: 0.8750
Epoch 4/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.3065 - accuracy: 0.8865 - val_loss: 0.3326 - val_accuracy: 0.8811
Epoch 5/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2880 - accuracy: 0.8930 - val_loss: 0.3208 - val_accuracy: 0.8843
Epoch 6/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.2744 - accuracy: 0.8981 - val_loss: 0.3313 - val_accuracy: 0.8810
Epoch 7/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2585 - accuracy: 0.9019 - val_loss: 0.3352 - val_accuracy: 0.8790
Epoch 8/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2445 - accuracy: 0.9078 - val_loss: 0.3151 - val_accuracy: 0.8849
Epoch 9/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.2366 - accuracy: 0.9113 - val_loss: 0.3167 - val_accuracy: 0.8881
Epoch 10/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.2241 - accuracy: 0.9162 - val_loss: 0.3258 - val_accuracy: 0.8857
Epoch 11/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.2158 - accuracy: 0.9194 - val_loss: 0.3087 - val_accuracy: 0.8927
Epoch 12/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.2091 - accuracy: 0.9218 - val_loss: 0.3287 - val_accuracy: 0.8904
Epoch 13/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1998 - accuracy: 0.9243 - val_loss: 0.3131 - val_accuracy: 0.8950
Epoch 14/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1937 - accuracy: 0.9271 - val_loss: 0.3177 - val_accuracy: 0.8925
Epoch 15/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1859 - accuracy: 0.9303 - val_loss: 0.3334 - val_accuracy: 0.8918
Epoch 16/50
1500/1500 [==============================] - 4s 2ms/step - loss: 0.1779 - accuracy: 0.9334 - val_loss: 0.3299 - val_accuracy: 0.8929
Epoch 17/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1743 - accuracy: 0.9348 - val_loss: 0.3391 - val_accuracy: 0.8920
Epoch 18/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1687 - accuracy: 0.9366 - val_loss: 0.3302 - val_accuracy: 0.8974
Epoch 19/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1628 - accuracy: 0.9385 - val_loss: 0.3641 - val_accuracy: 0.8868
Epoch 20/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1597 - accuracy: 0.9405 - val_loss: 0.3523 - val_accuracy: 0.8942
Epoch 21/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1534 - accuracy: 0.9434 - val_loss: 0.3584 - val_accuracy: 0.8951
Epoch 22/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1507 - accuracy: 0.9441 - val_loss: 0.3577 - val_accuracy: 0.8923
Epoch 23/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1453 - accuracy: 0.9452 - val_loss: 0.3807 - val_accuracy: 0.8957
Epoch 24/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1392 - accuracy: 0.9476 - val_loss: 0.3711 - val_accuracy: 0.8960
Epoch 25/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1364 - accuracy: 0.9494 - val_loss: 0.3731 - val_accuracy: 0.8940
Epoch 26/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1315 - accuracy: 0.9511 - val_loss: 0.3805 - val_accuracy: 0.8932
Epoch 27/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1319 - accuracy: 0.9507 - val_loss: 0.3966 - val_accuracy: 0.8880
Epoch 28/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1266 - accuracy: 0.9534 - val_loss: 0.3994 - val_accuracy: 0.8920
Epoch 29/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1207 - accuracy: 0.9546 - val_loss: 0.3918 - val_accuracy: 0.8959
Epoch 30/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1174 - accuracy: 0.9567 - val_loss: 0.4043 - val_accuracy: 0.8928
Epoch 31/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1191 - accuracy: 0.9546 - val_loss: 0.4114 - val_accuracy: 0.8951
Epoch 32/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1140 - accuracy: 0.9563 - val_loss: 0.4149 - val_accuracy: 0.8962
Epoch 33/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1121 - accuracy: 0.9574 - val_loss: 0.4373 - val_accuracy: 0.8931
Epoch 34/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1085 - accuracy: 0.9598 - val_loss: 0.4353 - val_accuracy: 0.8939
Epoch 35/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1056 - accuracy: 0.9591 - val_loss: 0.4325 - val_accuracy: 0.8938
Epoch 36/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1066 - accuracy: 0.9600 - val_loss: 0.4700 - val_accuracy: 0.8899
Epoch 37/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1019 - accuracy: 0.9615 - val_loss: 0.4440 - val_accuracy: 0.8947
Epoch 38/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0973 - accuracy: 0.9635 - val_loss: 0.4481 - val_accuracy: 0.8959
Epoch 39/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.1008 - accuracy: 0.9622 - val_loss: 0.4772 - val_accuracy: 0.8954
Epoch 40/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0919 - accuracy: 0.9653 - val_loss: 0.4723 - val_accuracy: 0.8916
Epoch 41/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0921 - accuracy: 0.9653 - val_loss: 0.4867 - val_accuracy: 0.8953
Epoch 42/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0919 - accuracy: 0.9657 - val_loss: 0.4710 - val_accuracy: 0.8936
Epoch 43/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0873 - accuracy: 0.9664 - val_loss: 0.4844 - val_accuracy: 0.8905
Epoch 44/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0884 - accuracy: 0.9669 - val_loss: 0.4972 - val_accuracy: 0.8963
Epoch 45/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0849 - accuracy: 0.9685 - val_loss: 0.4790 - val_accuracy: 0.8969
Epoch 46/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0831 - accuracy: 0.9687 - val_loss: 0.5028 - val_accuracy: 0.8945
Epoch 47/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0793 - accuracy: 0.9698 - val_loss: 0.5031 - val_accuracy: 0.8945
Epoch 48/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0806 - accuracy: 0.9693 - val_loss: 0.5065 - val_accuracy: 0.8990
Epoch 49/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0751 - accuracy: 0.9714 - val_loss: 0.5719 - val_accuracy: 0.8924
Epoch 50/50
1500/1500 [==============================] - 3s 2ms/step - loss: 0.0785 - accuracy: 0.9707 - val_loss: 0.5123 - val_accuracy: 0.8985
<keras.callbacks.History at 0x7fb39810a150>

Para finalizar este tutorial, avalie o hipermodelo nos dados de teste.

eval_result = hypermodel.evaluate(img_test, label_test)
print("[test loss, test accuracy]:", eval_result)
313/313 [==============================] - 1s 2ms/step - loss: 0.5632 - accuracy: 0.8908
[test loss, test accuracy]: [0.5631944537162781, 0.8907999992370605]

O diretório my_dir/intro_to_kt contém logs e pontos de verificação detalhados para cada tentativa (configuração de modelo) executada durante a pesquisa de hiperparâmetros. Se você executar novamente a pesquisa de hiperparâmetros, o Keras Tuner usará o estado existente desses logs para retomar a pesquisa. Para desabilitar esse comportamento, passe um argumento overwrite=True adicional ao instanciar o sintonizador.

Resumo

Neste tutorial, você aprendeu como usar o Keras Tuner para ajustar hiperparâmetros para um modelo. Para saber mais sobre o Keras Tuner, confira estes recursos adicionais:

Confira também o HParams Dashboard no TensorBoard para ajustar interativamente os hiperparâmetros do seu modelo.