How to make Keras compute a certain metric on validation data only?

Question

I'm using tf.keras with TensorFlow 1.14.0. I have implemented a custom metric that is quite computationally intensive and it slows down the training process if I simply add it to the list of metrics provided as model.compile(..., metrics=[...]).

How do I make Keras skip computation of the metric during training iterations but compute it on validation data (and print it) at the end of each epoch?

Answer 1

To do this you can create a tf.Variable in the metric calculation that determines if the calculation goes ahead and then update it when a test is run using a callback. e.g.

class MyCustomMetric(tf.keras.metrics.Metrics):

    def __init__(self, **kwargs):
        # Initialise as normal and add flag variable for when to run computation
        super(MyCustomMetric, self).__init__(**kwargs)
        self.metric_variable = self.add_weight(name='metric_varaible', initializer='zeros')
        self.update_metric = tf.Variable(False)

    def update_state(self, y_true, y_pred, sample_weight=None):
        # Use conditional to determine if computation is done
        if self.update_metric:
            # run computation
            self.metric_variable.assign_add(computation_result)

    def result(self):
        return self.metric_variable

    def reset_states(self):
        self.metric_variable.assign(0.)

class ToggleMetrics(tf.keras.callbacks.Callback):
    '''On test begin (i.e. when evaluate() is called or 
     validation data is run during fit()) toggle metric flag '''
    def on_test_begin(self, logs):
        for metric in self.model.metrics:
            if 'MyCustomMetric' in metric.name:
                metric.on.assign(True)
    def on_test_end(self,  logs):
        for metric in self.model.metrics:
            if 'MyCustomMetric' in metric.name:
                metric.on.assign(False)