def test_unweighted(self):
bce_obj = metrics.BinaryCrossentropy()
y_true = np.asarray([1, 0, 1, 0]).reshape([2, 2])
y_pred = np.asarray([1, 1, 1, 0], dtype=np.float32).reshape([2, 2])
result = bce_obj(y_true, y_pred)
assert np.allclose(K.eval(result), 3.833, atol=1e-3)
def test_weighted_from_logits(self):
bce_obj = metrics.BinaryCrossentropy(from_logits=True)
y_true = [[1, 0, 1], [0, 1, 1]]
y_pred = [[100.0, -100.0, 100.0], [100.0, 100.0, -100.0]]
sample_weight = [2., 2.5]
result = bce_obj(y_true, y_pred, sample_weight=sample_weight)
assert np.allclose(K.eval(result), 37.037, atol=1e-3)
def test_weighted(self):
bce_obj = metrics.BinaryCrossentropy()
y_true = np.asarray([1, 0, 1, 0]).reshape([2, 2])
y_pred = np.asarray([1, 1, 1, 0], dtype=np.float32).reshape([2, 2])
sample_weight = [1.5, 2.]
result = bce_obj(y_true, y_pred, sample_weight=sample_weight)
assert np.allclose(K.eval(result), 3.285, atol=1e-3)
def test_unweighted_with_logits(self):
bce_obj = metrics.BinaryCrossentropy(from_logits=True)
y_true = [[1, 0, 1], [0, 1, 1]]
y_pred = [[100.0, -100.0, 100.0], [100.0, 100.0, -100.0]]
result = bce_obj(y_true, y_pred)
assert np.allclose(K.eval(result), 33.333, atol=1e-3)
def test_config(self):
bce_obj = metrics.BinaryCrossentropy(
name='bce', dtype='int32', label_smoothing=0.2)
assert bce_obj.name == 'bce'
assert bce_obj.dtype == 'int32'
old_config = bce_obj.get_config()
assert np.allclose(old_config['label_smoothing'], 0.2, atol=1e-3)
def test_label_smoothing(self):
logits = ((100., -100., -100.))
y_true = ((1, 0, 1))
label_smoothing = 0.1
bce_obj = metrics.BinaryCrossentropy(
from_logits=True, label_smoothing=label_smoothing)
result = bce_obj(y_true, logits)
expected_value = (100.0 + 50.0 * label_smoothing) / 3.0
assert np.allclose(expected_value, K.eval(result), atol=1e-3)