def test_similar_vs_different_inputs(self):
loss = losses.PerceptualLoss()
pure_bright = tf.ones((3, 256, 256, 3)) * tf.constant([0.9, 0.7, 0.7])
pure_dark = tf.ones((3, 256, 256, 3)) * tf.constant([0.5, 0.2, 0.2])
speckles = tf.random.uniform((3, 256, 256, 3))
self.assertAllGreater(loss(pure_bright, speckles),
loss(pure_bright, pure_dark))
def test_composite_loss_equals_sum_of_components(self):
composite = losses.CompositeLoss()
mae = tf.keras.losses.MAE
vgg = losses.PerceptualLoss()
composite.add_loss(mae, 1.0)
composite.add_loss(vgg, 2.0)
y_true = tf.random.uniform((1, 192, 256, 3))
y_pred = tf.random.uniform((1, 192, 256, 3))
loss_value = composite(y_true, y_pred)
mae_loss_value = tf.math.reduce_mean(mae(y_true, y_pred))
vgg_loss_value = vgg(y_true, y_pred)
self.assertAllClose(loss_value,
mae_loss_value * 1.0 + vgg_loss_value * 2.0)
def test_different_inputs(self):
loss = losses.PerceptualLoss()
image_1 = tf.zeros((2, 192, 256, 3))
image_2 = tf.random.uniform((2, 192, 256, 3))
self.assertAllGreater(loss(image_1, image_2), 1.0)
def test_identical_inputs(self):
loss = losses.PerceptualLoss()
images = tf.random.uniform((2, 192, 256, 3))
self.assertAllClose(loss(images, images), 0.0)