Activation function for tensorflow DNNCalssifier's hidden layers

Question

It's possible to customize output layer's activation function for tensorflow's DNNClassifier canned estimator but no documentation about hidden layer's default activation function... Is it a ReLu one ? Is it customizable ?

Answer 1

You can modify the hidden layer's activation function by assigning new activation function to the activation_fn argument when you instantiate DNNClassifier.

classifier = tf.estimator.DNNClassifier(feature_columns=feature_columns,
                                        hidden_units=[10, 20, 10],
                                        activation_fn=tf.nn.sigmoid,  # default value is relu
                                        n_classes=3,
                                        model_dir="/tmp/model")