def test_masked_binary_crossentropy():
y_true = np.array([-1., -1, 0., 1.])
y_pred = np.array([1., 0., 0., 1.])
mbc = masked_binary_crossentropy(y_true, y_pred)
assert isinstance(mbc, tf.Tensor)
assert np.sum(mbc.numpy()) < 1e-5
def training_step(x, y):
with tf.GradientTape() as tape:
pred = model(x, training=True)
if focal_loss:
loss = masked_binary_focal_loss(y, pred, 2)
else:
loss = masked_binary_crossentropy(y, pred)
variables = model.trainable_variables
grads = tape.gradient(loss, variables)
opt.apply_gradients(zip(grads, variables))
return loss
def test_masked_focal_loss():
y_true = np.array([-1., -1, 0., 1.], dtype=np.float32)
y_pred = np.array([1., 0., 0.25, 0.75], dtype=np.float32)
mbc = masked_binary_crossentropy(y_true, y_pred)
focal_loss_gamma_0 = masked_binary_focal_loss(y_true, y_pred, 0)
focal_loss_gamma_2 = masked_binary_focal_loss(y_true, y_pred, 2)
assert isinstance(focal_loss_gamma_0, tf.Tensor)
# at gamma=0, focal loss collapses to crossentropy loss
assert abs(mbc.numpy() - focal_loss_gamma_0.numpy()) < 1e-5
# at higher gamma values the loss is pushed down
assert mbc.numpy() > focal_loss_gamma_2.numpy()
def loss(y_true, y_pred):
return masked_binary_crossentropy(
y_true, y_pred, label_smoothing=self.label_smoothing)