def evaluate(self):
"""Computes evaluation result.
Returns:
A named tuple with the following fields -
average_precision: a float number corresponding to average precision.
precisions: an array of precisions.
recalls: an array of recalls.
recall@50: recall computed on 50 top-scoring samples.
recall@100: recall computed on 100 top-scoring samples.
median_rank@50: median rank computed on 50 top-scoring samples.
median_rank@100: median rank computed on 100 top-scoring samples.
"""
if self._num_gt_instances == 0:
logging.warn('No ground truth instances')
if not self._scores:
scores = np.array([], dtype=float)
tp_fp_labels = np.array([], dtype=bool)
else:
scores = np.concatenate(self._scores)
tp_fp_labels = np.concatenate(self._tp_fp_labels)
relation_field_values = np.concatenate(self._relation_field_values)
for relation_field_value, _ in (
self._num_gt_instances_per_relationship.iteritems()):
precisions, recalls = metrics.compute_precision_recall(
scores[relation_field_values == relation_field_value],
tp_fp_labels[relation_field_values == relation_field_value],
self._num_gt_instances_per_relationship[relation_field_value])
self._average_precisions[
relation_field_value] = metrics.compute_average_precision(
precisions, recalls)
self._mean_average_precision = np.mean(self._average_precisions.values())
self._precisions, self._recalls = metrics.compute_precision_recall(
scores, tp_fp_labels, self._num_gt_instances)
self._weighted_average_precision = metrics.compute_average_precision(
self._precisions, self._recalls)
self._recall_50 = (
metrics.compute_recall_at_k(self._tp_fp_labels, self._num_gt_instances,
50))
self._median_rank_50 = (
metrics.compute_median_rank_at_k(self._tp_fp_labels, 50))
self._recall_100 = (
metrics.compute_recall_at_k(self._tp_fp_labels, self._num_gt_instances,
100))
self._median_rank_100 = (
metrics.compute_median_rank_at_k(self._tp_fp_labels, 100))
return VRDDetectionEvalMetrics(
self._weighted_average_precision, self._mean_average_precision,
self._average_precisions, self._precisions, self._recalls,
self._recall_50, self._recall_100, self._median_rank_50,
self._median_rank_100)
def evaluate(self):
"""Computes evaluation result.
Returns:
A named tuple with the following fields -
average_precision: a float number corresponding to average precision.
precisions: an array of precisions.
recalls: an array of recalls.
recall@50: recall computed on 50 top-scoring samples.
recall@100: recall computed on 100 top-scoring samples.
median_rank@50: median rank computed on 50 top-scoring samples.
median_rank@100: median rank computed on 100 top-scoring samples.
"""
if self._num_gt_instances == 0:
logging.warn('No ground truth instances')
if not self._scores:
scores = np.array([], dtype=float)
tp_fp_labels = np.array([], dtype=bool)
else:
scores = np.concatenate(self._scores)
tp_fp_labels = np.concatenate(self._tp_fp_labels)
relation_field_values = np.concatenate(self._relation_field_values)
for relation_field_value, _ in self._num_gt_instances_per_relationship.items(
):
precisions, recalls = metrics.compute_precision_recall(
scores[relation_field_values == relation_field_value],
tp_fp_labels[relation_field_values == relation_field_value],
self._num_gt_instances_per_relationship[relation_field_value])
self._average_precisions[
relation_field_value] = metrics.compute_average_precision(
precisions, recalls)
self._mean_average_precision = np.mean(
list(self._average_precisions.values()))
self._precisions, self._recalls = metrics.compute_precision_recall(
scores, tp_fp_labels, self._num_gt_instances)
self._weighted_average_precision = metrics.compute_average_precision(
self._precisions, self._recalls)
self._recall_50 = (metrics.compute_recall_at_k(self._tp_fp_labels,
self._num_gt_instances,
50))
self._median_rank_50 = (metrics.compute_median_rank_at_k(
self._tp_fp_labels, 50))
self._recall_100 = (metrics.compute_recall_at_k(
self._tp_fp_labels, self._num_gt_instances, 100))
self._median_rank_100 = (metrics.compute_median_rank_at_k(
self._tp_fp_labels, 100))
return VRDDetectionEvalMetrics(
self._weighted_average_precision, self._mean_average_precision,
self._average_precisions, self._precisions, self._recalls,
self._recall_50, self._recall_100, self._median_rank_50,
self._median_rank_100)
def test_compute_recall_at_k(self):
num_gt = 4
tp_fp = [
np.array([1, 0, 0], dtype=float),
np.array([0, 1], dtype=float),
np.array([0, 0, 0, 0, 0], dtype=float)
]
tp_fp_bool = [
np.array([True, False, False], dtype=bool),
np.array([False, True], dtype=float),
np.array([False, False, False, False, False], dtype=float)
]
recall_1 = metrics.compute_recall_at_k(tp_fp, num_gt, 1)
recall_3 = metrics.compute_recall_at_k(tp_fp, num_gt, 3)
recall_5 = metrics.compute_recall_at_k(tp_fp, num_gt, 5)
recall_3_bool = metrics.compute_recall_at_k(tp_fp_bool, num_gt, 3)
self.assertAlmostEqual(recall_1, 0.25)
self.assertAlmostEqual(recall_3, 0.5)
self.assertAlmostEqual(recall_3_bool, 0.5)
self.assertAlmostEqual(recall_5, 0.5)
def test_compute_recall_at_k(self):
num_gt = 4
tp_fp = [
np.array([1, 0, 0], dtype=float),
np.array([0, 1], dtype=float),
np.array([0, 0, 0, 0, 0], dtype=float)
]
tp_fp_bool = [
np.array([True, False, False], dtype=bool),
np.array([False, True], dtype=float),
np.array([False, False, False, False, False], dtype=float)
]
recall_1 = metrics.compute_recall_at_k(tp_fp, num_gt, 1)
recall_3 = metrics.compute_recall_at_k(tp_fp, num_gt, 3)
recall_5 = metrics.compute_recall_at_k(tp_fp, num_gt, 5)
recall_3_bool = metrics.compute_recall_at_k(tp_fp_bool, num_gt, 3)
self.assertAlmostEqual(recall_1, 0.25)
self.assertAlmostEqual(recall_3, 0.5)
self.assertAlmostEqual(recall_3_bool, 0.5)
self.assertAlmostEqual(recall_5, 0.5)