Refactor reset_state() in metrics.py to replace deprecated batch_set_value method.

[JacobJ17]

The reset_state function in coral_ordinal/metrics.py currently uses the tf.keras.backend.batch_set_value method, which has been deprecated in recent versions of TensorFlow/Keras.

File and Line Reference: reset_state function

def reset_state(self):
    """Resets all of the metric state variables at the start of each epoch."""
    K.batch_set_value([(v, 0) for v in self.variables])

I use coral-ordinal with Keras tuner, so the issue was caused after I started training by calling tuner.search. Training stopped after only 1 trial, and the only error message was AttributeError: 'str' object has no attribute 'name', however the program did not crash immediately. Instead, the program crashed a few lines later when calling tuner.get_best_models.

Proposed Solution: To avoid using the deprecated batch_set_value, I was able to use the .assign() method on each variable individually, as shown below:

def reset_state(self):
    for v in self.variables:
        v.assign(0)

This has worked for me with Tensorflow version 2.18.0 and Python 3.12.7