def _check_labels_and_predictions(predictions_dict, labels_dict):
"""Raise TypeError if the predictions and labels cannot be understood."""
if not (types.is_tensor(predictions_dict)
or prediction_keys.PredictionKeys.LOGISTIC in predictions_dict
or prediction_keys.PredictionKeys.PREDICTIONS in predictions_dict):
raise TypeError(
'cannot find predictions in %s. It is expected that either'
'predictions_dict is a tensor or it contains PredictionKeys.LOGISTIC'
'or PredictionKeys.PREDICTIONS.' % predictions_dict)
if not types.is_tensor(labels_dict):
raise TypeError('labels_dict is %s, which is not a tensor' %
labels_dict)
def is_compatible(self, features_dict, predictions_dict, labels_dict):
if (self._example_weight_key
and self._example_weight_key not in features_dict):
raise ValueError(
'example weight key %s not found in features_dict. '
'features were: %s' %
(self._example_weight_key, features_dict.keys()))
if (types.is_tensor(predictions_dict)
or prediction_keys.PredictionKeys.LOGISTIC in predictions_dict
or prediction_keys.PredictionKeys.PREDICTIONS
in predictions_dict):
if types.is_tensor(labels_dict):
return True
return False
def _get_prediction_tensor(predictions_dict):
"""Returns prediction Tensor for a specific Estimators.
Returns the prediction Tensor for some regression Estimators.
Args:
predictions_dict: Predictions dictionary.
Returns:
Predictions tensor, or None if none of the expected keys are found in
the predictions_dict.
"""
if types.is_tensor(predictions_dict):
return predictions_dict
key_precedence = (prediction_keys.PredictionKeys.LOGISTIC,
prediction_keys.PredictionKeys.PREDICTIONS,
prediction_keys.PredictionKeys.PROBABILITIES,
prediction_keys.PredictionKeys.LOGITS)
for key in key_precedence:
ref_tensor = predictions_dict.get(key)
if ref_tensor is not None:
return ref_tensor
return None
def _get_prediction_tensor(predictions_dict):
"""Returns prediction Tensor for a specific Estimators.
Returns the prediction Tensor for some regression Estimators.
Args:
predictions_dict: Predictions dictionary.
Returns:
Predictions tensor.
Raises:
KeyError: No expected keys are found in predictions_dict.
"""
if types.is_tensor(predictions_dict):
return predictions_dict
key_precedence = (prediction_keys.PredictionKeys.LOGISTIC,
prediction_keys.PredictionKeys.PREDICTIONS,
prediction_keys.PredictionKeys.PROBABILITIES,
prediction_keys.PredictionKeys.LOGITS)
for key in key_precedence:
ref_tensor = predictions_dict.get(key)
if ref_tensor is not None:
return ref_tensor
raise KeyError('cannot find any keys %s in predictions_dict %s.' %
(key_precedence, predictions_dict))
def _get_prediction_tensor(predictions_dict):
"""Returns prediction Tensor for a specific Estimators.
Returns the prediction Tensor for some regression Estimators.
Args:
predictions_dict: Predictions dictionary.
Returns:
Predictions tensor.
"""
if types.is_tensor(predictions_dict):
ref_tensor = predictions_dict
else:
ref_tensor = predictions_dict.get(
prediction_keys.PredictionKeys.LOGISTIC)
if ref_tensor is None:
ref_tensor = predictions_dict.get(
prediction_keys.PredictionKeys.PREDICTIONS)
return ref_tensor
def check_compatibility(self, features_dict, predictions_dict,
labels_dict):
if not isinstance(predictions_dict, dict):
raise TypeError(
'predictions_dict should be a dict. predictions_dict '
'was: %s' % predictions_dict)
if prediction_keys.PredictionKeys.CLASSES not in predictions_dict:
raise KeyError(
'predictions_dict should contain PredictionKeys.CLASSES. '
'predictions_dict was: %s' % predictions_dict)
if prediction_keys.PredictionKeys.PROBABILITIES not in predictions_dict:
raise KeyError(
'predictions_dict should contain '
'PredictionKeys.PROBABILITIES. predictions_dict was: %s' %
predictions_dict)
if not types.is_tensor(labels_dict):
raise TypeError(
'labels_dict should be a tensor. labels_dict was: %s' %
labels_dict)
_check_weight_present(features_dict, self._example_weight_key)
def get_metric_ops(self, features_dict, predictions_dict, labels_dict):
ref_tensor = _get_prediction_tensor(predictions_dict)
if ref_tensor is None:
# Note that if predictions_dict is a Tensor and not a dict,
# get_predictions_tensor will return predictions_dict, so if we get
# here, if means that predictions_dict is a dict without any of the
# standard keys.
#
# If we can't get any of standard keys, then pick the first key
# in alphabetical order if the predictions dict is non-empty.
# If the predictions dict is empty, try the labels dict.
# If that is empty too, default to the empty Tensor.
tf.logging.info(
'ExampleCount post export metric: could not find any of '
'the standard keys in predictions_dict (keys were: %s)',
predictions_dict.keys())
if predictions_dict is not None and predictions_dict.keys():
first_key = sorted(predictions_dict.keys())[0]
ref_tensor = predictions_dict[first_key]
tf.logging.info(
'Using the first key from predictions_dict: %s', first_key)
elif labels_dict is not None:
if types.is_tensor(labels_dict):
ref_tensor = labels_dict
tf.logging.info('Using the labels Tensor')
elif labels_dict.keys():
first_key = sorted(labels_dict.keys())[0]
ref_tensor = labels_dict[first_key]
tf.logging.info('Using the first key from labels_dict: %s',
first_key)
if ref_tensor is None:
tf.logging.info(
'Could not find a reference Tensor for example count. '
'Defaulting to the empty Tensor.')
ref_tensor = tf.constant([])
return {
metric_keys.EXAMPLE_COUNT: metrics.total(tf.shape(ref_tensor)[0])
}
def is_compatible(self, features_dict, predictions_dict, labels_dict):
if (types.is_tensor(predictions_dict)
or prediction_keys.PredictionKeys.LOGISTIC in predictions_dict
or prediction_keys.PredictionKeys.PREDICTIONS
in predictions_dict):
return True
def _check_labels(labels_dict):
"""Raise TypeError if the labels cannot be understood."""
if not types.is_tensor(labels_dict):
raise TypeError('labels_dict is %s, which is not a tensor' %
labels_dict)
