docs/ipynb/image_regression.ipynb
{
"cells": [
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"!pip install autokeras"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"import tensorflow as tf\n",
"from keras.datasets import mnist\n",
"\n",
"import autokeras as ak"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"To make this tutorial easy to follow, we just treat MNIST dataset as a\n",
"regression dataset. It means we will treat prediction targets of MNIST dataset,\n",
"which are integers ranging from 0 to 9 as numerical values, so that they can be\n",
"directly used as the regression targets.\n",
"\n",
"## A Simple Example\n",
"The first step is to prepare your data. Here we use the MNIST dataset as an\n",
"example\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"(x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
"x_train = x_train[:100]\n",
"y_train = y_train[:100]\n",
"x_test = x_test[:100]\n",
"y_test = y_test[:100]\n",
"print(x_train.shape) # (60000, 28, 28)\n",
"print(y_train.shape) # (60000,)\n",
"print(y_train[:3]) # array([7, 2, 1], dtype=uint8)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"The second step is to run the ImageRegressor. It is recommended have more\n",
"trials for more complicated datasets. This is just a quick demo of MNIST, so\n",
"we set max_trials to 1. For the same reason, we set epochs to 1. You can also\n",
"leave the epochs unspecified for an adaptive number of epochs.\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"# Initialize the image regressor.\n",
"reg = ak.ImageRegressor(overwrite=True, max_trials=1)\n",
"# Feed the image regressor with training data.\n",
"reg.fit(x_train, y_train, epochs=1)\n",
"\n",
"\n",
"# Predict with the best model.\n",
"predicted_y = reg.predict(x_test)\n",
"print(predicted_y)\n",
"\n",
"\n",
"# Evaluate the best model with testing data.\n",
"print(reg.evaluate(x_test, y_test))"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"## Validation Data\n",
"By default, AutoKeras use the last 20% of training data as validation data. As\n",
"shown in the example below, you can use validation_split to specify the\n",
"percentage.\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"reg.fit(\n",
" x_train,\n",
" y_train,\n",
" # Split the training data and use the last 15% as validation data.\n",
" validation_split=0.15,\n",
" epochs=1,\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"You can also use your own validation set instead of splitting it from the\n",
"training data with validation_data.\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"split = 50000\n",
"x_val = x_train[split:]\n",
"y_val = y_train[split:]\n",
"x_train = x_train[:split]\n",
"y_train = y_train[:split]\n",
"reg.fit(\n",
" x_train,\n",
" y_train,\n",
" # Use your own validation set.\n",
" validation_data=(x_val, y_val),\n",
" epochs=2,\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"## Customized Search Space\n",
"For advanced users, you may customize your search space by using AutoModel\n",
"instead of ImageRegressor. You can configure the ImageBlock for some high-level\n",
"configurations, e.g., block_type for the type of neural network to search,\n",
"normalize for whether to do data normalization, augment for whether to do data\n",
"augmentation. You can also do not specify these arguments, which would leave\n",
"the different choices to be tuned automatically. See the following example for\n",
"detail.\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"input_node = ak.ImageInput()\n",
"output_node = ak.ImageBlock(\n",
" # Only search ResNet architectures.\n",
" block_type=\"resnet\",\n",
" # Normalize the dataset.\n",
" normalize=False,\n",
" # Do not do data augmentation.\n",
" augment=False,\n",
")(input_node)\n",
"output_node = ak.RegressionHead()(output_node)\n",
"reg = ak.AutoModel(\n",
" inputs=input_node, outputs=output_node, overwrite=True, max_trials=1\n",
")\n",
"reg.fit(x_train, y_train, epochs=1)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"The usage of AutoModel is similar to the functional API of Keras. Basically,\n",
"you are building a graph, whose edges are blocks and the nodes are intermediate\n",
"outputs of blocks. To add an edge from input_node to output_node with\n",
"output_node = ak.[some_block]([block_args])(input_node).\n",
"\n",
"You can even also use more fine grained blocks to customize the search space\n",
"even further. See the following example.\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"input_node = ak.ImageInput()\n",
"output_node = ak.Normalization()(input_node)\n",
"output_node = ak.ImageAugmentation(horizontal_flip=False)(output_node)\n",
"output_node = ak.ResNetBlock(version=\"v2\")(output_node)\n",
"output_node = ak.RegressionHead()(output_node)\n",
"reg = ak.AutoModel(\n",
" inputs=input_node, outputs=output_node, overwrite=True, max_trials=1\n",
")\n",
"reg.fit(x_train, y_train, epochs=1)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"## Data Format\n",
"The AutoKeras ImageRegressor is quite flexible for the data format.\n",
"\n",
"For the image, it accepts data formats both with and without the channel\n",
"dimension. The images in the MNIST dataset do not have the channel dimension.\n",
"Each image is a matrix with shape (28, 28). AutoKeras also accepts images of\n",
"three dimensions with the channel dimension at last, e.g., (32, 32, 3), (28,\n",
"28, 1).\n",
"\n",
"For the regression targets, it should be a vector of numerical values.\n",
"AutoKeras accepts numpy.ndarray.\n",
"\n",
"We also support using tf.data.Dataset format for the training data. In this\n",
"case, the images would have to be 3-dimentional.\n"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"(x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
"x_train = x_train[:100]\n",
"y_train = y_train[:100]\n",
"x_test = x_test[:100]\n",
"y_test = y_test[:100]\n",
"\n",
"# Reshape the images to have the channel dimension.\n",
"x_train = x_train.reshape(x_train.shape + (1,))\n",
"x_test = x_test.reshape(x_test.shape + (1,))\n",
"y_train = y_train.reshape(y_train.shape + (1,))\n",
"y_test = y_test.reshape(y_test.shape + (1,))\n",
"\n",
"print(x_train.shape) # (60000, 28, 28, 1)\n",
"print(y_train.shape) # (60000, 10)\n",
"\n",
"train_set = tf.data.Dataset.from_tensor_slices(((x_train,), (y_train,)))\n",
"test_set = tf.data.Dataset.from_tensor_slices(((x_test,), (y_test,)))\n",
"\n",
"reg = ak.ImageRegressor(overwrite=True, max_trials=1)\n",
"# Feed the tensorflow Dataset to the regressor.\n",
"reg.fit(train_set, epochs=1)\n",
"# Predict with the best model.\n",
"predicted_y = reg.predict(test_set)\n",
"# Evaluate the best model with testing data.\n",
"print(reg.evaluate(test_set))"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"## Reference\n",
"[ImageRegressor](/image_regressor),\n",
"[AutoModel](/auto_model/#automodel-class),\n",
"[ImageBlock](/block/#imageblock-class),\n",
"[Normalization](/block/#normalization-class),\n",
"[ImageAugmentation](/block/#image-augmentation-class),\n",
"[ResNetBlock](/block/#resnetblock-class),\n",
"[ImageInput](/node/#imageinput-class),\n",
"[RegressionHead](/block/#regressionhead-class).\n"
]
}
],
"metadata": {
"colab": {
"collapsed_sections": [],
"name": "image_regression",
"private_outputs": false,
"provenance": [],
"toc_visible": true
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.0"
}
},
"nbformat": 4,
"nbformat_minor": 0
}