def _setup_metrics(self):
self.metric_functions[LOSS] = self.eval_loss_function
self.metric_functions[ERROR] = ErrorScore(name='metric_error')
self.metric_functions[MEAN_SQUARED_ERROR] = MeanSquaredErrorMetric(
name='metric_mse')
self.metric_functions[MEAN_ABSOLUTE_ERROR] = MeanAbsoluteErrorMetric(
name='metric_mae')
self.metric_functions[R2] = R2Score(name='metric_r2')
def _setup_metrics(self):
self.metric_functions = {} # needed to shadow class variable
if self.loss[TYPE] == 'mean_squared_error':
self.metric_functions[LOSS] = MSEMetric(name='eval_loss')
else:
self.metric_functions[LOSS] = MAEMetric(name='eval_loss')
self.metric_functions[ERROR] = ErrorScore(name='metric_error')
self.metric_functions[MEAN_SQUARED_ERROR] = MeanSquaredErrorMetric(
name='metric_mse')
self.metric_functions[MEAN_ABSOLUTE_ERROR] = MeanAbsoluteErrorMetric(
name='metric_mae')
self.metric_functions[R2] = R2Score(name='metric_r2')
def test_ErrorScore(generated_data):
error_score = ErrorScore()
assert np.isnan(error_score.result().numpy())
# test as single batch
error_score.update_state(generated_data.y_true, generated_data.y_good)
good_single_batch = error_score.result().numpy()
assert np.isreal(good_single_batch)
# test as two batches
error_score.reset_states()
error_score.update_state(generated_data.y_true[:SPLIT_POINT],
generated_data.y_good[:SPLIT_POINT])
error_score.update_state(generated_data.y_true[SPLIT_POINT:],
generated_data.y_good[SPLIT_POINT:])
good_two_batch = error_score.result().numpy()
assert np.isreal(good_two_batch)
# single batch and multi-batch should be very close
assert np.isclose(good_single_batch, good_two_batch)
# test for bad predictions
error_score.reset_states()
error_score.update_state(generated_data.y_true[:SPLIT_POINT],
generated_data.y_bad[:SPLIT_POINT])
error_score.update_state(generated_data.y_true[SPLIT_POINT:],
generated_data.y_bad[SPLIT_POINT:])
bad_prediction_score = error_score.result().numpy()
# magnitude of bad predictions should be greater than good predictions
assert np.abs(bad_prediction_score) > np.abs(good_two_batch)
