def test_should_return_higher_value_for_less_frequent_occuring_class(self):
with tf.Graph().as_default():
with tf.Session():
frequencies = [2, 1]
result = tf_calculate_efnet_weights_for_frequency_by_label(
frequencies).eval()
assert result[0] < result[1]
def test_should_return_zero_value_for_not_occuring_class(self):
with tf.Graph().as_default():
with tf.Session():
frequencies = [1, 0]
result = tf_calculate_efnet_weights_for_frequency_by_label(
frequencies).eval()
assert result[-1] == 0.0
def test_should_return_same_value_for_classes_with_same_frequencies(self):
with tf.Graph().as_default():
with tf.Session():
frequencies = [1, 1]
result = tf_calculate_efnet_weights_for_frequency_by_label(
frequencies).eval()
assert result[0] == result[1]
def _create_pos_weights_tensor(
base_loss,
separate_channel_annotation_tensor,
pos_weight_values,
input_uri,
debug):
frequency_by_label = tf.reduce_sum(
separate_channel_annotation_tensor,
axis=[0, 1],
keep_dims=True,
name='frequency_by_channel'
)
pos_weight_sample = tf_calculate_efnet_weights_for_frequency_by_label(
frequency_by_label
)
pos_weight = (
pos_weight_sample * pos_weight_values
if base_loss == BaseLoss.WEIGHTED_SAMPLE_WEIGHTED_CROSS_ENTROPY
else pos_weight_sample
)
if debug:
pos_weight = tf.Print(
pos_weight, [
pos_weight,
pos_weight_sample,
frequency_by_label,
input_uri
],
'pos weights, sample, frequency, uri: ',
summarize=1000
)
get_logger().debug(
'pos_weight before batch: %s (frequency_by_label: %s)',
pos_weight, frequency_by_label
)
return pos_weight