def test_supported_loss_returns_correctly_with_loss_kwargs(self):
import haiku as hk
def net_function(x: jnp.ndarray) -> jnp.ndarray:
net = hk.Sequential([])
return net(x)
net_transform = hk.transform(net_function)
actual_loss_function_wrapper = get_haiku_loss_function(
net_transform,
loss="sigmoid_focal_crossentropy",
alpha=None,
gamma=None)
# Check works
rng = jax.random.PRNGKey(42)
params = net_transform.init(rng, jnp.array(0))
self.assertEqual(
0,
actual_loss_function_wrapper(params,
x=jnp.array([0]),
y_true=jnp.array([0])))
self.assertEqual(
0,
actual_loss_function_wrapper(params,
x=jnp.array([1]),
y_true=jnp.array([1])))
# When alpha and gamma are None, it should be equal to log_loss
self.assertEqual(
log_loss(y_true=jnp.array([0.27]), y_pred=jnp.array([0])),
actual_loss_function_wrapper(params,
x=jnp.array([0]),
y_true=jnp.array([0.27])),
)
self.assertEqual(
log_loss(y_true=jnp.array([0.97]), y_pred=jnp.array([1])),
actual_loss_function_wrapper(params,
x=jnp.array([1]),
y_true=jnp.array([0.97])),
)
def test_multiclass_returns_correctly(self):
actual_loss = log_loss(jnp.array([[0, 1], [1, 0]]),
jnp.array([[0, 1], [1, 0]]))
self.assertEqual(0, actual_loss)
actual_loss = log_loss(jnp.array([[0, 1], [1, 0]]),
jnp.array([[0, 1], [1, 0]]),
normalize=False)
np.testing.assert_array_equal(jnp.array([0, 0]), actual_loss)
# Based on scikit-learn: https://github.com/scikit-learn/scikit-learn/blob
# /ffbb1b4a0bbb58fdca34a30856c6f7faace87c67/sklearn/metrics/tests/test_classification.py#L2135
actual_loss = log_loss(
jnp.array([[0, 1, 0], [1, 0, 0], [0, 0, 1]]),
jnp.array([[0.2, 0.7, 0.1], [0.6, 0.2, 0.2], [0.6, 0.1, 0.3]]))
self.assertEqual(0.69049114, actual_loss)
actual_loss = log_loss(
jnp.array([[0, 1, 0], [1, 0, 0], [0, 0, 1]]),
jnp.array([[0.2, 0.7, 0.1], [0.6, 0.2, 0.2], [0.6, 0.1, 0.3]]),
normalize=False,
)
np.testing.assert_array_equal(
jnp.array([0.35667497, 0.5108256, 1.2039728]), actual_loss)
def test_binary_returns_correctly(self):
actual_loss = sigmoid_focal_crossentropy(jnp.array([1]),
jnp.array([1]))
self.assertEqual(0, actual_loss)
actual_loss = sigmoid_focal_crossentropy(jnp.array([0]),
jnp.array([0]))
self.assertEqual(0, actual_loss)
actual_loss = sigmoid_focal_crossentropy(jnp.array([1, 0]),
jnp.array([1, 0]))
self.assertEqual(0, actual_loss)
# Based on tensorflow_addons: https://github.com/tensorflow/addons/blob/v0.10.0/tensorflow_addons/losses
# /tests/focal_loss_test.py#L106
actual_loss = sigmoid_focal_crossentropy(
jnp.array([1, 1, 1, 0, 0, 0]),
jnp.array([0.97, 0.91, 0.73, 0.27, 0.09, 0.03]),
alpha=None,
gamma=None)
# When alpha and gamma are None, it should be equal to log_loss
expected_loss = log_loss(
jnp.array([1, 1, 1, 0, 0, 0]),
jnp.array([0.97, 0.91, 0.73, 0.27, 0.09, 0.03]))
self.assertEqual(expected_loss, actual_loss)
actual_loss = sigmoid_focal_crossentropy(
jnp.array([1, 1, 1, 0, 0, 0]),
jnp.array([0.97, 0.91, 0.73, 0.27, 0.09, 0.03]),
alpha=None,
gamma=2.0)
self.assertEqual(0.007911247, actual_loss)
actual_loss = sigmoid_focal_crossentropy(
jnp.array([1, 1, 1, 0, 0, 0]),
jnp.array([0.97, 0.91, 0.73, 0.27, 0.09, 0.03]),
alpha=None,
gamma=2.0,
normalize=False,
)
np.testing.assert_array_equal(
jnp.array([
2.7413207e-05, 7.6391589e-04, 2.2942409e-02, 2.2942409e-02,
7.6391734e-04, 2.7413207e-05
]),
actual_loss,
)
def test_binary_returns_correctly(self):
actual_loss = log_loss(jnp.array([1]), jnp.array([1]))
self.assertEqual(0, actual_loss)
actual_loss = log_loss(jnp.array([0]), jnp.array([0]))
self.assertEqual(0, actual_loss)
actual_loss = log_loss(jnp.array([1, 0]), jnp.array([1, 0]))
self.assertEqual(0, actual_loss)
actual_loss = log_loss(jnp.array([1, 0]), jnp.array([1, 1]))
self.assertEqual(17.269388, actual_loss)
# Based on scikit-learn: https://github.com/scikit-learn/scikit-learn/blob
# /ffbb1b4a0bbb58fdca34a30856c6f7faace87c67/sklearn/metrics/tests/test_classification.py#L2135
actual_loss = log_loss(jnp.array([0, 0, 0, 1, 1, 1]),
jnp.array([0.5, 0.9, 0.99, 0.1, 0.25, 0.999]))
self.assertEqual(1.8817972, actual_loss)
actual_loss = log_loss(jnp.array([0, 0, 0, 1, 1, 1]),
jnp.array([0.5, 0.9, 0.99, 0.1, 0.25, 0.999]),
normalize=False)
np.testing.assert_array_equal(
jnp.array([
6.9314718e-01, 2.3025849e00, 4.6051712e00, 2.3025851e00,
1.3862944e00, 1.0004875e-03
]),
actual_loss,
)
def test_raises_when_attempt_to_use_not_one_hot_encoded_multiclass(self):
with self.assertRaises(TypeError) as _:
log_loss(
jnp.array([1, 0, 2]),
jnp.array([[0.2, 0.7, 0.1], [0.6, 0.2, 0.2], [0.6, 0.1, 0.3]]))
def test_raises_when_number_of_multiclass_classes_not_equal(self):
with self.assertRaises(TypeError) as _:
log_loss(jnp.array([[0, 0, 1], [0, 0, 1], [0, 0, 1]]),
jnp.array([[0.2, 0.7], [0.6, 0.5], [0.4, 0.1]]))
def test_raises_when_number_of_samples_not_equal_multiclass(self):
with self.assertRaises(ValueError) as _:
log_loss(jnp.array([[0, 1], [1, 0]]),
jnp.array([[0.2, 0.7], [0.6, 0.5], [0.4, 0.1]]))
def test_raises_when_number_of_samples_not_equal(self):
with self.assertRaises(ValueError) as _:
log_loss(jnp.array([0, 1]), jnp.array([0, 1, 0]))