How to name custom metrics in Keras fit output

Question

I am using custom metrics when training my Keras model. It works fine, except that the metrics names in the output of model.fit_generator(...) are not interpretable (NB: Tensorboard also uses these wrong names).

Here is a reproducible example of what I am doing: the metrics are using one parameter (in addition to the prediction and ground truth), so I defined a factory to generate a parameterless metric function, similar to this:

def my_dummy_metric(y_true, y_pred, the_param=1.0):
    return the_param * keras.backend.ones((1))

def my_metric_factory(the_param=1.0):
    def fn(y_true, y_pred):
        return my_dummy_metric(y_true, y_pred, the_param=the_param)

    return fn

my_second_metric = my_metric_factory(2.0)
my_other_metric = my_metric_factory(3.14)

Then I compile and train my model:

model.compile(my_optim, my_loss, [my_second_metric, my_other_metric])
history = model.fit_generator(...)
print(history.params['metrics'])

My trouble is that the metric names in history are fn, fn_1, val_fn and val_fn_1. These names are also used by Tensorboard and you need to know the implem details to understand them.

On the contrary, I don't have this problem when using a simple custom function, without factory:

model.compile(my_optim, my_loss, [my_dummy_metric])
history = model.fit_generator(...)
print(history.params['metrics'])

Would it be possible to obtain my_XXX_metric as output names also in the factory-based use case?

Environment: Using Keras 2.2.4, TF 1.14.0, Python 3.7

Answer 1

Yes, this is possible. In the metric factory, just set an appropriate __name__ of metric function. For example:

def my_metric_factory(the_param=1.0):
    def fn(y_true, y_pred):
        return my_dummy_metric(y_true, y_pred, the_param=the_param)

    fn.__name__ = 'metricname_{}'.format(the_param)
    return fn