Skip to content

Image Classification

View in Colab     GitHub source

!pip install autokeras

A Simple Example

The first step is to prepare your data. Here we use the MNIST dataset as an example

# One-hot encode the labels.

import numpy as np
import tensorflow as tf
from tensorflow.keras.datasets import mnist
from tensorflow.python.keras.utils.data_utils import Sequence
import autokeras as ak
(x_train, y_train), (x_test, y_test) = mnist.load_data()
print(x_train.shape)  # (60000, 28, 28)
print(y_train.shape)  # (60000,)
print(y_train[:3])  # array([7, 2, 1], dtype=uint8)

The second step is to run the ImageClassifier.

# Initialize the image classifier.
clf = ak.ImageClassifier(max_trials=10)  # It tries 10 different models.
# Feed the image classifier with training data., y_train, epochs=3)

# Predict with the best model.
predicted_y = clf.predict(x_test)

# Evaluate the best model with testing data.
print(clf.evaluate(x_test, y_test))

Validation Data

By default, AutoKeras use the last 20% of training data as validation data. As shown in the example below, you can use validation_split to specify the percentage.
    # Split the training data and use the last 15% as validation data.

You can also use your own validation set instead of splitting it from the training data with validation_data.

split = 50000
x_val = x_train[split:]
y_val = y_train[split:]
x_train = x_train[:split]
y_train = y_train[:split]
    # Use your own validation set.
    validation_data=(x_val, y_val),

Customized Search Space

For advanced users, you may customize your search space by using AutoModel instead of ImageClassifier. You can configure the ImageBlock for some high-level configurations, e.g., block_type for the type of neural network to search, normalize for whether to do data normalization, augment for whether to do data augmentation. You can also do not specify these arguments, which would leave the different choices to be tuned automatically. See the following example for detail.

input_node = ak.ImageInput()
output_node = ak.ImageBlock(
    # Only search ResNet architectures.
    # Normalize the dataset.
    # Do not do data augmentation.
output_node = ak.ClassificationHead()(output_node)
clf = ak.AutoModel(inputs=input_node, outputs=output_node, max_trials=10), y_train, epochs=3)

The usage of AutoModel is similar to the functional API of Keras. Basically, you are building a graph, whose edges are blocks and the nodes are intermediate outputs of blocks. To add an edge from input_node to output_node with output_node = ak.some_block(input_node).

You can even also use more fine grained blocks to customize the search space even further. See the following example.

input_node = ak.ImageInput()
output_node = ak.Normalization()(input_node)
output_node = ak.ImageAugmentation(percentage=0.3)(output_node)
output_node = ak.ResNetBlock(version="v2")(output_node)
output_node = ak.ClassificationHead()(output_node)
clf = ak.AutoModel(inputs=input_node, outputs=output_node, max_trials=10), y_train, epochs=3)

Data Format

The AutoKeras ImageClassifier is quite flexible for the data format.

For the image, it accepts data formats both with and without the channel dimension. The images in the MNIST dataset do not have the channel dimension. Each image is a matrix with shape (28, 28). AutoKeras also accepts images of three dimensions with the channel dimension at last, e.g., (32, 32, 3), (28, 28, 1).

For the classification labels, AutoKeras accepts both plain labels, i.e. strings or integers, and one-hot encoded encoded labels, i.e. vectors of 0s and 1s.

So if you prepare your data in the following way, the ImageClassifier should still work.

(x_train, y_train), (x_test, y_test) = mnist.load_data()

# Reshape the images to have the channel dimension.
x_train = x_train.reshape(x_train.shape + (1,))
x_test = x_test.reshape(x_test.shape + (1,))

eye = np.eye(10)
y_train = eye[y_train]
y_test = eye[y_test]

print(x_train.shape)  # (60000, 28, 28, 1)
print(y_train.shape)  # (60000, 10)
# array([[0., 0., 0., 0., 0., 1., 0., 0., 0., 0.],
#        [1., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
#        [0., 0., 0., 0., 1., 0., 0., 0., 0., 0.]])

We also support using format for the training data. In this case, the images would have to be 3-dimentional. The labels have to be one-hot encoded for multi-class classification to be wrapped into tensorflow Dataset.

train_set =,), (y_train,)))
test_set =,), (y_test,)))

clf = ak.ImageClassifier(max_trials=10)
# Feed the tensorflow Dataset to the classifier.
# Predict with the best model.
predicted_y = clf.predict(test_set)
# Evaluate the best model with testing data.


ImageClassifier, AutoModel, ImageBlock, Normalization, ImageAugmentation, ResNetBlock, ImageInput, ClassificationHead.