def _find_label_domain(
eval_config: config_pb2.EvalConfig, schema: schema_pb2.Schema,
model_name: Text, output_name: Text
) -> Tuple[Optional[Union[int, float]], Optional[Union[int, float]]]:
"""Find the min and max value for the label_key for this model / output."""
model_spec = model_util.get_model_spec(eval_config, model_name)
if not model_spec:
return None, None
label_key = model_util.get_label_key(model_spec, output_name)
if not label_key:
return None, None
label_schema = None
for feature_schema in schema.feature:
if feature_schema.name == label_key:
label_schema = feature_schema
break
if label_schema is None:
return None, None
# Find the domain
if label_schema.HasField('int_domain'):
label_domain = label_schema.int_domain
elif label_schema.HasField('float_domain'):
label_domain = label_schema.float_domain
else:
return None, None
left, right = None, None
if label_domain.HasField('min'):
left = float(label_domain.min)
if label_domain.HasField('max'):
right = float(label_domain.max)
return left, right
def _create_accumulator(
self) -> tf_metric_accumulators.TFCompilableMetricsAccumulator:
padding_options = None
if self._eval_config is not None:
model_spec = model_util.get_model_spec(self._eval_config,
self._model_name)
if model_spec is not None and model_spec.HasField(
'padding_options'):
padding_options = model_spec.padding_options
return tf_metric_accumulators.TFCompilableMetricsAccumulator(
padding_options,
self._output_counts,
desired_batch_size=self._desired_batch_size)
def create_accumulator(
self) -> tf_metric_accumulators.TFCompilableMetricsAccumulator:
configs = zip(self._metric_configs, self._loss_configs)
padding_options = None
if self._eval_config is not None:
model_spec = model_util.get_model_spec(self._eval_config,
self._model_name)
if model_spec is not None and model_spec.HasField('padding_options'):
padding_options = model_spec.padding_options
return tf_metric_accumulators.TFCompilableMetricsAccumulator(
padding_options, [len(m) + len(l) for m, l in configs],
desired_batch_size=self._desired_batch_size)
def _get_example_weights(self, model_name: str, features: Dict[str,
Any]) -> Any:
spec = model_util.get_model_spec(self._eval_config, model_name)
if not spec:
raise ValueError(
'Missing model_spec for model_name "{}"'.format(model_name))
if spec.example_weight_key:
if spec.example_weight_key not in features:
raise ValueError(
'Missing feature for example_weight_key "{}": features={}'.format(
spec.example_weight_key, features))
return features[spec.example_weight_key]
elif spec.example_weight_keys:
example_weights = {}
for k, v in spec.example_weight_keys.items():
if v not in features:
raise ValueError(
'Missing feature for example_weight_key "{}": features={}'.format(
k, features))
example_weights[k] = features[v]
return example_weights
else:
return np.array([1.0])