# Tensorflow & Keras¶

Since keras is a higher-lever interface for tensorflow and nowadays part of tensorflow , we do not need to distinguish between keras and tensorflow models when using ceml.

Computing a counterfactual of a tensorflow/keras model is done by using the ceml.tfkeras.counterfactual.generate_counterfactual() function.

Note

We have to run in eager execution mode when computing a counterfactual! Since tensorflow 2, eager execution is enabled by default.

We must provide the tensorflow/keras model within a class that is derived from the ceml.model.model.ModelWithLoss class. In this class, we must overwrite the predict function and get_loss function which returns a loss that we want to use - a couple of differentiable loss functions are implemented in ceml.backend.tensorflow.costfunctions.

Besides the model, we must specify the input whose prediction we want to explain and the desired target prediction (prediction of the counterfactual). In addition we can restrict the features that can be used for computing a counterfactual, specify a regularization of the counterfactual and specifying the optimization algorithm used for computing a counterfactual.

A complete example of a softmax regression model using the negative-log-likelihood is given below:

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 #!/usr/bin/env python3 # -*- coding: utf-8 -*- import tensorflow as tf import numpy as np from sklearn.datasets import load_iris from sklearn.model_selection import train_test_split from sklearn.metrics import accuracy_score from ceml.tfkeras import generate_counterfactual from ceml.backend.tensorflow.costfunctions import NegLogLikelihoodCost from ceml.model import ModelWithLoss # Neural network - Softmax regression class Model(ModelWithLoss): def __init__(self, input_size, num_classes): super(Model, self).__init__() self.model = tf.keras.models.Sequential([ tf.keras.layers.Dense(num_classes, activation='softmax', input_shape=(input_size,)) ]) def fit(self, x_train, y_train, num_epochs=800): self.model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) self.model.fit(x_train, y_train, epochs=num_epochs, verbose=False) def predict(self, x): return np.argmax(self.model(x), axis=1) def predict_proba(self, x): return self.model(x) def __call__(self, x): return self.predict(x) def get_loss(self, y_target, pred=None): return NegLogLikelihoodCost(input_to_output=self.model.predict_proba, y_target=y_target) if __name__ == "__main__": tf.random.set_seed(42) # Fix random seed # Load data X, y = load_iris(return_X_y=True) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=1) # Create and fit model model = Model(4, 3) model.fit(X_train, y_train) # Evaluation y_pred = model.predict(X_test) print("Accuracy: {0}".format(accuracy_score(y_test, y_pred))) # Select a data point whose prediction has to be explained x_orig = X_test[1,:] print("Prediction on x: {0}".format(model.predict(np.array([x_orig])))) # Whitelist of features we can use/change when computing the counterfactual features_whitelist = None # Compute counterfactual optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate=1.0) # Init optimization algorithm optimizer_args = {"max_iter": 1000} print("\nCompute counterfactual ....") print(generate_counterfactual(model, x_orig, y_target=0, features_whitelist=features_whitelist, regularization="l1", C=0.01, optimizer=optimizer, optimizer_args=optimizer_args))