hosa.models.cnn.cnn_models.CNNClassification

class hosa.models.cnn.cnn_models.CNNClassification(n_outputs, n_kernels, n_neurons_dense_layer=50, optimizer='adam', metrics=None, cnn_dim=1, kernel_size=2, pool_size=2, strides_convolution=1, strides_pooling=2, padding='valid', dropout_percentage=0.1, activation_function_gol='relu', activation_function_dense='relu', batch_size=1000, epochs=50, patience=5, **kwargs)[source]

Bases: hosa.models.cnn.cnn_models.BaseCNN

Convolutional Neural Network (CNN) classifier.

The model comprises an input layer, a set of groups of layers (GofLayers [Men20])—where each group is composed of one convolution layer, followed by one pooling layer and a dropout layer—, a dense layer, and an output layer. The output layer is a dense layer with n_outputs units, with the softmax activation function.

Note

The parameters used in this library were adapted from the exact parameters of the TensorFlow library. Descriptions were thus modified accordingly to our approach. However, refer to the TensorFlow documentation for more details about each of those parameters.

Parameters

  • n_outputs (int) – Number of class labels to predict.

  • n_kernels (list) – i-th element represents the number of output filters of the convolution layer in the i-th GofLayer.

  • n_neurons_dense_layer (int) – Number of neuron units in the dense layer (i.e., the dense layer after the set of groups of layers).

  • optimizer (str) – Name of the optimizer. See tensorflow.keras.optimizers.

  • metrics (list) – List of metrics to be evaluated by the model during training and testing. Each item of the list can be a string (name of a TensorFlow’s built-in function), function, or a tf.keras.metrics.Metric instance. If None, metrics will default to ['accuracy'].

  • cnn_dim (int) – Number of dimensions applicable to all the convolution layers of the GofLayers.

  • kernel_size (int or tuple) – Integer or tuple/list of integers, specifying the length (for cnn_dim = 1), the height and width (for cnn_dim = 2), or the depth, height and width (for cnn_dim = 3) of the convolution window. It can also be a single integer to specify the same value for all spatial dimensions. This applies to all the convolution layers of the GofLayers.

  • pool_size (int or tuple) – Size of the max-pooling window applicable to all the max-pooling layers of the GofLayers. For cnn_dim = 1, use a tuple with 1 integer; For cnn_dim = 2, a tuple with 2 integers and for cnn_dim = 3, a tuple with 3 integers. It can also be a single integer to specify the same value for all spatial dimensions.

  • strides_convolution (int or tuple) – Integer or tuple/list of integers, specifying the strides of the convolution. For cnn_dim = 1, use a tuple with 1 integer; For cnn_dim = 2, a tuple with 2 integers and for cnn_dim = 3, a tuple with 3 integers. It can also be a single integer to specify the same value for all spatial dimensions. This applies to all the convolution layers of the GofLayers.

  • strides_pooling (int or tuple) – Integer or tuple/list of integers, specifying the strides of the pooling. For cnn_dim = 1, use a tuple with 1 integer; For cnn_dim = 2, a tuple with 2 integers and for cnn_dim = 3, a tuple with 3 integers. It can also be a single integer to specify the same value for all spatial dimensions. This applies to all the pooling layers of the GofLayers.

  • padding (str) – Available options are valid or same. valid means no padding. same results in padding evenly to the left/right or up/down of the input such that output has the same height/width dimension as the input. This applies to all the pooling layers of the GofLayers.

  • dropout_percentage (float) – Fraction of the input units to drop. This applies to all the dropout layers of the GofLayers.

  • activation_function_gol (str or None) – Activation function to use in the convolution layers of the GofLayers. If not specified, no activation is applied (i.e., uses the linear activation function). See tensorflow.keras.activations. This applies to all the convolution layers of the GofLayers.

  • activation_function_dense (str or None) –

    Activation function to use on the dense layer (i.e., the dense layer after the set of groups of layers). If not specified, no activation is applied (i.e., uses the linear activation function). See tensorflow.keras.activations.

  • batch_size (int or None) – Number of samples per batch of computation. If None, batch_size will default to 32.

  • epochs (int) – Maximum number of epochs to train the model.

  • patience (int) – Number of epochs with no improvement after which training will be stopped.

  • **kwargs – Ignored. Extra arguments that are used for compatibility’s sake.

Examples

 1from tensorflow import keras
 2
 3from hosa.models.cnn import CNNClassification
 4
 5# 1 - Load and split the data
 6fashion_mnist = keras.datasets.fashion_mnist
 7(x_train, y_train), (X_test, y_test) = fashion_mnist.load_data()
 8# 2 - Normalize the images, and reshape the images for cnn input
 9x_train = x_train / 255.0
10X_test = X_test / 255.0
11x_train = x_train.reshape((-1, 28 * 28))
12X_test = X_test.reshape((-1, 28 * 28))
13# 3 - Create and fit the model
14clf = CNNClassification(10, 10, [4, 3])
15clf.prepare(x_train, y_train)
16clf.compile()
17clf.fit(x_train, y_train)
18# 4 - Calculate predictions
19clf.predict(X_test)
20# 5 - Compute the score
21score = clf.score(X_test, y_test)

Methods

add_gol(n_kernel, cnn_dim)

Adds a GofLayer to the estimator.

aux_fit(x, y, callback, validation_size[, ...])

Auxiliar function for classification and regression models compatibility.

compile()

Compiles the model for training.

fit(x, y[, validation_size, shuffle, atol, ...])

Fits the model to data matrix x and target(s) y.

predict(x, **kwargs)

Predicts the target values using the input data in the trained model.

prepare(x, y)

Prepares the model by adding the layers to the estimator: input layer, GofLayers, flatten and dense layers.

score(x, y[, imbalance_correction])

Computes the performance metrics on the given input data and target values.