 View on TensorFlow.org
|
 Run in Google Colab
|
 View source on GitHub
|
 Download notebook
|
Overview
The Keras Tuner is a library that helps you pick the optimal set of hyperparameters for your TensorFlow program. The process of selecting the right set of hyperparameters for your machine learning (ML) application is called hyperparameter tuning or hypertuning.
Hyperparameters are the variables that govern the training process and the topology of an ML model. These variables remain constant over the training process and directly impact the performance of your ML program. Hyperparameters are of two types:
- Model hyperparameters which influence model selection such as the number and width of hidden layers
- Algorithm hyperparameters which influence the speed and quality of the learning algorithm such as the learning rate for Stochastic Gradient Descent (SGD) and the number of nearest neighbors for a k Nearest Neighbors (KNN) classifier
In this tutorial, you will use the Keras Tuner to perform hypertuning for an image classification application.
Setup
import tensorflow as tf
from tensorflow import keras
Install and import the Keras Tuner.
pip install -q -U keras-tuner
import keras_tuner as kt
Download and prepare the dataset
In this tutorial, you will use the Keras Tuner to find the best hyperparameters for a machine learning model that classifies images of clothing from the Fashion MNIST dataset.
Load the data.
(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
Define the model
When you build a model for hypertuning, you also define the hyperparameter search space in addition to the model architecture. The model you set up for hypertuning is called a hypermodel.
You can define a hypermodel through two approaches:
- By using a model builder function
- By subclassing the
HyperModelclass of the Keras Tuner API
You can also use two pre-defined HyperModel classes - HyperXception and HyperResNet for computer vision applications.
In this tutorial, you use a model builder function to define the image classification model. The model builder function returns a compiled model and uses hyperparameters you define inline to hypertune the model.
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
Instantiate the tuner and perform hypertuning
Instantiate the tuner to perform the hypertuning. The Keras Tuner has four tuners available - RandomSearch, Hyperband, BayesianOptimization, and Sklearn. In this tutorial, you use the Hyperband tuner.
To instantiate the Hyperband tuner, you must specify the hypermodel, the objective to optimize and the maximum number of epochs to train (max_epochs).
tuner = kt.Hyperband(model_builder,
objective='val_accuracy',
max_epochs=10,
factor=3,
directory='my_dir',
project_name='intro_to_kt')
The Hyperband tuning algorithm uses adaptive resource allocation and early-stopping to quickly converge on a high-performing model. This is done using a sports championship style bracket. The algorithm trains a large number of models for a few epochs and carries forward only the top-performing half of models to the next round. Hyperband determines the number of models to train in a bracket by computing 1 + logfactor(max_epochs) and rounding it up to the nearest integer.
Create a callback to stop training early after reaching a certain value for the validation loss.
stop_early = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=5)
Run the hyperparameter search. The arguments for the search method are the same as those used for tf.keras.model.fit in addition to the callback above.
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 41s] val_accuracy: 0.8865000009536743 Best val_accuracy So Far: 0.8865000009536743 Total elapsed time: 00h 08m 39s INFO:tensorflow:Oracle triggered exit The hyperparameter search is complete. The optimal number of units in the first densely-connected layer is 416 and the optimal learning rate for the optimizer is 0.001.
Train the model
Find the optimal number of epochs to train the model with the hyperparameters obtained from the search.
# 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 3ms/step - loss: 0.4962 - accuracy: 0.8246 - val_loss: 0.4271 - val_accuracy: 0.8409 Epoch 2/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3665 - accuracy: 0.8673 - val_loss: 0.3703 - val_accuracy: 0.8627 Epoch 3/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3290 - accuracy: 0.8791 - val_loss: 0.3278 - val_accuracy: 0.8832 Epoch 4/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3040 - accuracy: 0.8875 - val_loss: 0.3544 - val_accuracy: 0.8710 Epoch 5/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2843 - accuracy: 0.8938 - val_loss: 0.3508 - val_accuracy: 0.8742 Epoch 6/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2724 - accuracy: 0.8986 - val_loss: 0.3314 - val_accuracy: 0.8813 Epoch 7/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2575 - accuracy: 0.9042 - val_loss: 0.3174 - val_accuracy: 0.8871 Epoch 8/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2466 - accuracy: 0.9075 - val_loss: 0.3292 - val_accuracy: 0.8833 Epoch 9/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2345 - accuracy: 0.9124 - val_loss: 0.3428 - val_accuracy: 0.8864 Epoch 10/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2241 - accuracy: 0.9159 - val_loss: 0.3126 - val_accuracy: 0.8918 Epoch 11/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2149 - accuracy: 0.9186 - val_loss: 0.3332 - val_accuracy: 0.8851 Epoch 12/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2107 - accuracy: 0.9199 - val_loss: 0.3087 - val_accuracy: 0.8939 Epoch 13/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2003 - accuracy: 0.9245 - val_loss: 0.3184 - val_accuracy: 0.8931 Epoch 14/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1934 - accuracy: 0.9272 - val_loss: 0.3197 - val_accuracy: 0.8946 Epoch 15/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1886 - accuracy: 0.9294 - val_loss: 0.3224 - val_accuracy: 0.8912 Epoch 16/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1784 - accuracy: 0.9339 - val_loss: 0.3417 - val_accuracy: 0.8921 Epoch 17/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1754 - accuracy: 0.9343 - val_loss: 0.3287 - val_accuracy: 0.8947 Epoch 18/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1683 - accuracy: 0.9363 - val_loss: 0.3427 - val_accuracy: 0.8926 Epoch 19/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1622 - accuracy: 0.9394 - val_loss: 0.3369 - val_accuracy: 0.8943 Epoch 20/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1570 - accuracy: 0.9410 - val_loss: 0.3449 - val_accuracy: 0.8907 Epoch 21/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1550 - accuracy: 0.9414 - val_loss: 0.3443 - val_accuracy: 0.8929 Epoch 22/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1501 - accuracy: 0.9442 - val_loss: 0.3449 - val_accuracy: 0.8979 Epoch 23/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1440 - accuracy: 0.9457 - val_loss: 0.3613 - val_accuracy: 0.8939 Epoch 24/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1382 - accuracy: 0.9484 - val_loss: 0.3676 - val_accuracy: 0.8894 Epoch 25/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1369 - accuracy: 0.9477 - val_loss: 0.3940 - val_accuracy: 0.8871 Epoch 26/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1305 - accuracy: 0.9512 - val_loss: 0.3630 - val_accuracy: 0.8935 Epoch 27/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1251 - accuracy: 0.9520 - val_loss: 0.4069 - val_accuracy: 0.8840 Epoch 28/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1263 - accuracy: 0.9520 - val_loss: 0.3820 - val_accuracy: 0.8928 Epoch 29/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1214 - accuracy: 0.9542 - val_loss: 0.3970 - val_accuracy: 0.8955 Epoch 30/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1174 - accuracy: 0.9554 - val_loss: 0.3996 - val_accuracy: 0.8895 Epoch 31/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1143 - accuracy: 0.9567 - val_loss: 0.4196 - val_accuracy: 0.8923 Epoch 32/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1132 - accuracy: 0.9574 - val_loss: 0.4209 - val_accuracy: 0.8902 Epoch 33/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1089 - accuracy: 0.9592 - val_loss: 0.4093 - val_accuracy: 0.8951 Epoch 34/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1085 - accuracy: 0.9597 - val_loss: 0.4049 - val_accuracy: 0.8988 Epoch 35/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1016 - accuracy: 0.9611 - val_loss: 0.4444 - val_accuracy: 0.8907 Epoch 36/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1035 - accuracy: 0.9613 - val_loss: 0.4821 - val_accuracy: 0.8899 Epoch 37/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1009 - accuracy: 0.9618 - val_loss: 0.4409 - val_accuracy: 0.8896 Epoch 38/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0952 - accuracy: 0.9623 - val_loss: 0.4678 - val_accuracy: 0.8915 Epoch 39/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0928 - accuracy: 0.9653 - val_loss: 0.4495 - val_accuracy: 0.8939 Epoch 40/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0967 - accuracy: 0.9645 - val_loss: 0.4850 - val_accuracy: 0.8902 Epoch 41/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0927 - accuracy: 0.9642 - val_loss: 0.4785 - val_accuracy: 0.8904 Epoch 42/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0874 - accuracy: 0.9678 - val_loss: 0.4565 - val_accuracy: 0.8949 Epoch 43/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0825 - accuracy: 0.9689 - val_loss: 0.5582 - val_accuracy: 0.8848 Epoch 44/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0841 - accuracy: 0.9686 - val_loss: 0.5025 - val_accuracy: 0.8915 Epoch 45/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0826 - accuracy: 0.9695 - val_loss: 0.5150 - val_accuracy: 0.8856 Epoch 46/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0817 - accuracy: 0.9695 - val_loss: 0.5074 - val_accuracy: 0.8943 Epoch 47/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0789 - accuracy: 0.9699 - val_loss: 0.5018 - val_accuracy: 0.8947 Epoch 48/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0796 - accuracy: 0.9701 - val_loss: 0.5014 - val_accuracy: 0.8921 Epoch 49/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0774 - accuracy: 0.9711 - val_loss: 0.4936 - val_accuracy: 0.8959 Epoch 50/50 1500/1500 [==============================] - 4s 3ms/step - loss: 0.0787 - accuracy: 0.9709 - val_loss: 0.5335 - val_accuracy: 0.8897 Best epoch: 34
Re-instantiate the hypermodel and train it with the optimal number of epochs from above.
hypermodel = tuner.hypermodel.build(best_hps)
# Retrain the model
hypermodel.fit(img_train, label_train, epochs=best_epoch, validation_split=0.2)
Epoch 1/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.4928 - accuracy: 0.8247 - val_loss: 0.4264 - val_accuracy: 0.8497 Epoch 2/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3677 - accuracy: 0.8659 - val_loss: 0.3683 - val_accuracy: 0.8650 Epoch 3/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3316 - accuracy: 0.8780 - val_loss: 0.3367 - val_accuracy: 0.8779 Epoch 4/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3047 - accuracy: 0.8866 - val_loss: 0.3295 - val_accuracy: 0.8798 Epoch 5/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2851 - accuracy: 0.8942 - val_loss: 0.3325 - val_accuracy: 0.8794 Epoch 6/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2727 - accuracy: 0.8974 - val_loss: 0.3389 - val_accuracy: 0.8819 Epoch 7/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2550 - accuracy: 0.9046 - val_loss: 0.3226 - val_accuracy: 0.8867 Epoch 8/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2465 - accuracy: 0.9066 - val_loss: 0.3186 - val_accuracy: 0.8871 Epoch 9/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2354 - accuracy: 0.9121 - val_loss: 0.3306 - val_accuracy: 0.8883 Epoch 10/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2258 - accuracy: 0.9158 - val_loss: 0.3254 - val_accuracy: 0.8870 Epoch 11/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2172 - accuracy: 0.9176 - val_loss: 0.3299 - val_accuracy: 0.8857 Epoch 12/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2081 - accuracy: 0.9209 - val_loss: 0.3135 - val_accuracy: 0.8916 Epoch 13/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2014 - accuracy: 0.9248 - val_loss: 0.3375 - val_accuracy: 0.8892 Epoch 14/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1930 - accuracy: 0.9275 - val_loss: 0.3354 - val_accuracy: 0.8912 Epoch 15/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1878 - accuracy: 0.9293 - val_loss: 0.3193 - val_accuracy: 0.8948 Epoch 16/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1808 - accuracy: 0.9321 - val_loss: 0.3391 - val_accuracy: 0.8901 Epoch 17/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1756 - accuracy: 0.9339 - val_loss: 0.3466 - val_accuracy: 0.8877 Epoch 18/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1709 - accuracy: 0.9349 - val_loss: 0.3345 - val_accuracy: 0.8910 Epoch 19/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1624 - accuracy: 0.9389 - val_loss: 0.3441 - val_accuracy: 0.8938 Epoch 20/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1564 - accuracy: 0.9415 - val_loss: 0.3611 - val_accuracy: 0.8885 Epoch 21/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1536 - accuracy: 0.9424 - val_loss: 0.3552 - val_accuracy: 0.8934 Epoch 22/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1470 - accuracy: 0.9455 - val_loss: 0.3595 - val_accuracy: 0.8952 Epoch 23/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1436 - accuracy: 0.9459 - val_loss: 0.3675 - val_accuracy: 0.8936 Epoch 24/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1402 - accuracy: 0.9468 - val_loss: 0.3706 - val_accuracy: 0.8922 Epoch 25/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1371 - accuracy: 0.9492 - val_loss: 0.3785 - val_accuracy: 0.8933 Epoch 26/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1297 - accuracy: 0.9509 - val_loss: 0.3624 - val_accuracy: 0.8942 Epoch 27/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1315 - accuracy: 0.9507 - val_loss: 0.3854 - val_accuracy: 0.8953 Epoch 28/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1247 - accuracy: 0.9523 - val_loss: 0.3988 - val_accuracy: 0.8900 Epoch 29/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1216 - accuracy: 0.9532 - val_loss: 0.4011 - val_accuracy: 0.8886 Epoch 30/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1189 - accuracy: 0.9554 - val_loss: 0.3949 - val_accuracy: 0.8907 Epoch 31/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1153 - accuracy: 0.9568 - val_loss: 0.3940 - val_accuracy: 0.8963 Epoch 32/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1112 - accuracy: 0.9575 - val_loss: 0.3962 - val_accuracy: 0.8947 Epoch 33/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1102 - accuracy: 0.9586 - val_loss: 0.4067 - val_accuracy: 0.8956 Epoch 34/34 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1050 - accuracy: 0.9605 - val_loss: 0.4112 - val_accuracy: 0.8926 <keras.callbacks.History at 0x7fc6b60ce2d0>
To finish this tutorial, evaluate the hypermodel on the test data.
eval_result = hypermodel.evaluate(img_test, label_test)
print("[test loss, test accuracy]:", eval_result)
313/313 [==============================] - 1s 2ms/step - loss: 0.4548 - accuracy: 0.8854 [test loss, test accuracy]: [0.4547567367553711, 0.8853999972343445]
The my_dir/intro_to_kt directory contains detailed logs and checkpoints for every trial (model configuration) run during the hyperparameter search. If you re-run the hyperparameter search, the Keras Tuner uses the existing state from these logs to resume the search. To disable this behavior, pass an additional overwrite=True argument while instantiating the tuner.
Summary
In this tutorial, you learned how to use the Keras Tuner to tune hyperparameters for a model. To learn more about the Keras Tuner, check out these additional resources:
Also check out the HParams Dashboard in TensorBoard to interactively tune your model hyperparameters.