def test_gelu(self, dtype):
x = tf.constant([-2.0, -1.0, 0.0, 1.0, 2.0], dtype=dtype)
expected_result = tf.constant(
[-0.04540229, -0.158808, 0.0, 0.841192, 1.9545977], dtype=dtype)
self.assertAllCloseAccordingToType(gelu(x), expected_result)
expected_result = tf.constant(
[-0.04550028, -0.15865526, 0.0, 0.8413447, 1.9544997], dtype=dtype)
self.assertAllCloseAccordingToType(gelu(x, False), expected_result)
def test_gelu(dtype):
x = tf.constant([-2.0, -1.0, 0.0, 1.0, 2.0], dtype=dtype)
expected_result = tf.constant(
[-0.04540229, -0.158808, 0.0, 0.841192, 1.9545977], dtype=dtype)
test_utils.assert_allclose_according_to_type(gelu(x), expected_result)
expected_result = tf.constant(
[-0.04550028, -0.15865526, 0.0, 0.8413447, 1.9544997], dtype=dtype)
test_utils.assert_allclose_according_to_type(gelu(x, False),
expected_result)
def test_unknown_shape(self):
fn = gelu.get_concrete_function(
tf.TensorSpec(shape=None, dtype=tf.float32))
for shape in [(1, ), (1, 2), (1, 2, 3), (1, 2, 3, 4)]:
x = tf.ones(shape=shape, dtype=tf.float32)
self.assertAllClose(fn(x), gelu(x))
def test_theoretical_gradients(dtype, approximate):
# Only test theoretical gradients for float32 and float64
# because of the instability of float16 while computing jacobian
x = tf.constant([-2.0, -1.0, 0.0, 1.0, 2.0], dtype=dtype)
theoretical, numerical = tf.test.compute_gradient(
lambda x: gelu(x, approximate=approximate), [x]
)
test_utils.assert_allclose_according_to_type(theoretical, numerical, atol=1e-4)
def test_theoretical_gradients(self, dtype):
# Only test theoretical gradients for float32 and float64
# because of the instability of float16 while computing jacobian
x = tf.constant([-2.0, -1.0, 0.0, 1.0, 2.0], dtype=dtype)
for approximate in [True, False]:
with self.subTest(approximate=approximate):
theoretical, numerical = tf.test.compute_gradient(
lambda x: gelu(x, approximate=approximate), [x]
)
self.assertAllCloseAccordingToType(theoretical, numerical, atol=1e-4)
def test_gradients(self, dtype):
x = tf.constant([1.0, 2.0, 3.0], dtype=dtype)
for approximate in [True, False]:
with self.subTest(approximate=approximate):
with tf.GradientTape(persistent=True) as tape:
tape.watch(x)
y_ref = _ref_gelu(x, approximate)
y = gelu(x, approximate)
grad_ref = tape.gradient(y_ref, x)
grad = tape.gradient(y, x)
self.assertAllCloseAccordingToType(grad, grad_ref)
def verify_funcs_are_equivalent(self, dtype):
x_np = np.random.uniform(-10, 10, size=(4, 4)).astype(dtype)
x = tf.convert_to_tensor(x_np)
for approximate in [True, False]:
with tf.GradientTape(persistent=True) as t:
t.watch(x)
y_native = gelu(x, approximate=approximate)
y_py = _gelu_py(x, approximate=approximate)
self.assertAllCloseAccordingToType(y_native, y_py, atol=1e-4)
grad_native = t.gradient(y_native, x)
grad_py = t.gradient(y_py, x)
self.assertAllCloseAccordingToType(grad_native, grad_py, atol=1e-4)
def verify_funcs_are_equivalent(dtype, approximate):
x_np = np.random.uniform(-10, 10, size=(4, 4)).astype(dtype)
x = tf.convert_to_tensor(x_np)
with tf.GradientTape(persistent=True) as t:
t.watch(x)
y_native = gelu(x, approximate=approximate)
y_py = _gelu_py(x, approximate=approximate)
test_utils.assert_allclose_according_to_type(y_native, y_py)
grad_native = t.gradient(y_native, x)
grad_py = t.gradient(y_py, x)
# TODO: lower atol to 1e-6
# currently it doesn't work.
# It necessitates changing the Python or C++ implementation.
test_utils.assert_allclose_according_to_type(grad_native, grad_py, atol=1e-5)
def call(self, inputs):
return gelu(inputs, approximate=self.approximate)
def test_gelu(self, dtype):
x = np.random.rand(2, 3, 4).astype(dtype)
self.assertAllCloseAccordingToType(gelu(x), _ref_gelu(x))
self.assertAllCloseAccordingToType(gelu(x, False), _ref_gelu(x, False))
