def _evaluate(self, dataset):
if dataset is None:
logger.info("No validation dataset is configured")
return
eval_metrics = EvaluationMetrics(self.eval_metrics_fn())
for features, labels in dataset:
outputs = self.model_inst.call(features)
if not isinstance(outputs, dict):
outputs = {MetricsDictKey.MODEL_OUTPUT: outputs}
eval_metrics.update_evaluation_metrics(outputs, labels)
metrics = eval_metrics.get_evaluation_summary()
logger.info("Evaluation metrics : {}".format(metrics))
return metrics
class EvaluationJob(object):
"""Representation of an evaluation job"""
def __init__(self, metrics_dict, model_version, total_tasks=-1):
"""
Args:
metrics_dict: A python dictionary. If model has only one output,
`metrics_dict` is a dictionary of `{metric_name: metric}`,
i.e. `{"acc": tf.keras.metrics.Accuracy()}`.
If model has multiple outputs, `metric_dict` is a dictionary of
`{output_name: {metric_name: metric}}`,
i.e. `{
"output_a": {"acc": tf.keras.metrics.Accuracy()},
"output_b": {"auc": tf.keras.metrics.AUC()},
}`. Note that for model with multiple outputs, each metric
only uses one output.
model_version: The version of the model to be evaluated.
total_tasks: The number of evaluation tasks.
"""
self.model_version = model_version
self._total_tasks = total_tasks
self._completed_tasks = 0
self.evaluation_metrics = EvaluationMetrics(metrics_dict)
def complete_task(self):
self._completed_tasks += 1
def finished(self):
return self._completed_tasks >= self._total_tasks
def report_evaluation_metrics(self, model_outputs_pb, labels):
labels = pb_to_ndarray(labels)
model_outputs = {}
for name, tensor_pb in model_outputs_pb.items():
model_outputs[name] = pb_to_ndarray(tensor_pb)
self.evaluation_metrics.update_evaluation_metrics(
model_outputs, labels
)