def test_zero_weighted(self):
cl_obj = contrastive.ContrastiveLoss()
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5],
dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred, sample_weight=0.0)
self.assertAllClose(loss, 0.0)
def test_zero_loss():
cl_obj = contrastive.ContrastiveLoss()
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred)
np.testing.assert_allclose(loss, 0.0)
def test_no_reduction(self):
cl_obj = contrastive.ContrastiveLoss(reduction=tf.keras.losses.Reduction.NONE)
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5], dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred)
# Loss = y * (y`)^2 + (1 - y) * (max(m - y`, 0))^2
# = [max(1 - 0.1, 0)^2, max(1 - 0.3, 0)^2,
# 1.3^2, 0.7^2, max(1 - 1.1, 0)^2, 0.5^2]
# = [0.9^2, 0.7^2, 1.3^2, 0.7^2, 0^2, 0.5^2]
# = [0.81, 0.49, 1.69, 0.49, 0, 0.25]
self.assertAllClose(loss, [0.81, 0.49, 1.69, 0.49, 0.0, 0.25])
def test_unweighted(self):
cl_obj = contrastive.ContrastiveLoss()
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5], dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred)
# Loss = y * (y`)^2 + (1 - y) * (max(m - y`, 0))^2
# = [max(1 - 0.1, 0)^2, max(1 - 0.3, 0)^2,
# 1.3^2, 0.7^2, max(1 - 1.1, 0)^2, 0.5^2]
# = [0.9^2, 0.7^2, 1.3^2, 0.7^2, 0^2, 0.5^2]
# = [0.81, 0.49, 1.69, 0.49, 0, 0.25]
# Reduced loss = (0.81 + 0.49 + 1.69 + 0.49 + 0 + 0.25) / 6
# = 0.621666
self.assertAllClose(loss, 0.621666)
def test_non_default_margin(self):
cl_obj = contrastive.ContrastiveLoss(margin=2.0)
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5], dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred)
# Loss = y * (y`)^2 + (1 - y) * (max(m - y`, 0))^2
# = [max(2 - 0.1, 0)^2, max(2 - 0.3, 0)^2,
# 1.3^2, 0.7^2, max(2 - 1.1, 0)^2, 0.5^2]
# = [1.9^2, 1.7^2, 1.3^2, 0.7^2, 0.9^2, 0.5^2]
# = [3.61, 2.89, 1.69, 0.49, 0.81, 0.25]
# Reduced loss = (3.61 + 2.89 + 1.69 + 0.49 + 0.81 + 0.25) / 6
# = 1.623333
loss = self.evaluate(loss)
self.assertAllClose(loss, 1.623333)
def test_scalar_weighted(self):
cl_obj = contrastive.ContrastiveLoss()
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5], dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred, sample_weight=6.0)
# Loss = y * (y`)^2 + (1 - y) * (max(m - y`, 0))^2
# = [max(1 - 0.1, 0)^2, max(1 - 0.3, 0)^2,
# 1.3^2, 0.7^2, max(1 - 1.1, 0)^2, 0.5^2]
# = [0.9^2, 0.7^2, 1.3^2, 0.7^2, 0^2, 0.5^2]
# = [0.81, 0.49, 1.69, 0.49, 0, 0.25]
# Weighted loss = [0.81 * 6, 0.49 * 6, 1.69 * 6,
# 0.49 * 6, 0 * 6, 0.25 * 6]
# Reduced loss = (0.81 + 0.49 + 1.69 + 0.49 + 0 + 0.25) * 6 / 6
# = 3.73
self.assertAllClose(loss, 3.73)
def test_sum_reduction():
cl_obj = contrastive.ContrastiveLoss(
reduction=tf.keras.losses.Reduction.SUM)
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5],
dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred)
# Loss = y * (y`)^2 + (1 - y) * (max(m - y`, 0))^2
# = [max(1 - 0.1, 0)^2, max(1 - 0.3, 0)^2,
# 1.3^2, 0.7^2, max(1 - 1.1, 0)^2, 0.5^2]
# = [0.9^2, 0.7^2, 1.3^2, 0.7^2, 0^2, 0.5^2]
# = [0.81, 0.49, 1.69, 0.49, 0, 0.25]
# Reduced loss = 0.81 + 0.49 + 1.69 + 0.49 + 0 + 0.25
# = 3.73
np.testing.assert_allclose(loss, 3.73)
def test_sample_weighted(self):
cl_obj = contrastive.ContrastiveLoss()
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([0.1, 0.3, 1.3, 0.7, 1.1, 0.5], dtype=tf.dtypes.float32)
sample_weight = tf.constant(
[1.2, 0.8, 0.5, 0.4, 1.5, 1.0], dtype=tf.dtypes.float32
)
loss = cl_obj(y_true, y_pred, sample_weight=sample_weight)
# Loss = y * (y`)^2 + (1 - y) * (max(m - y`, 0))^2
# = [max(1 - 0.1, 0)^2, max(1 - 0.3, 0)^2,
# 1.3^2, 0.7^2, max(1 - 1.1, 0)^2, 0.5^2]
# = [0.9^2, 0.7^2, 1.3^2, 0.7^2, 0^2, 0.5^2]
# = [0.81, 0.49, 1.69, 0.49, 0, 0.25]
# Weighted loss = [0.81 * 1.2, 0.49 * 0.8, 1.69 * 0.5,
# 0.49 * 0.4, 0 * 1.5, 0.25 * 1.0]
# = [0.972, 0.392, 0.845, 0.196, 0, 0.25]
# Reduced loss = (0.972 + 0.392 + 0.845 + 0.196 + 0 + 0.25) / 6
# = 0.4425
self.assertAllClose(loss, 0.4425)
def test_config(self):
cl_obj = contrastive.ContrastiveLoss(
reduction=tf.keras.losses.Reduction.SUM, name="cl"
)
self.assertEqual(cl_obj.name, "cl")
self.assertEqual(cl_obj.reduction, tf.keras.losses.Reduction.SUM)
def test_zero_loss(self):
cl_obj = contrastive.ContrastiveLoss()
y_true = tf.constant([0, 0, 1, 1, 0, 1], dtype=tf.dtypes.int64)
y_pred = tf.constant([1., 1., 0., 0., 1., 0.], dtype=tf.dtypes.float32)
loss = cl_obj(y_true, y_pred)
self.assertAllClose(loss, 0.0)
def test_config():
cl_obj = contrastive.ContrastiveLoss(
reduction=tf.keras.losses.Reduction.SUM, name="cl")
assert cl_obj.name == "cl"
assert cl_obj.reduction == tf.keras.losses.Reduction.SUM
