def testGradientWithRhsIsZero(self):
x_vals = [0, 1.0, np.nan, np.inf, np.NINF]
x = constant_op.constant(x_vals, dtype=dtypes.float32)
y = array_ops.zeros_like(x, dtype=dtypes.float32)
outputs = math_ops.mul_no_nan(x, y)
with self.cached_session():
dx, dy = gradients.gradients(outputs, [x, y])
self.assertAllClose(dx.eval(), np.zeros(x.shape.as_list()))
self.assertAllClose(dy.eval(), x_vals)
def testBasicGradient(self):
inputs = constant_op.constant(np.arange(-3, 3), dtype=dtypes.float32)
outputs = math_ops.mul_no_nan(inputs, 1 + math_ops.abs(inputs))
with self.cached_session():
error = gradient_checker.compute_gradient_error(
inputs,
inputs.get_shape().as_list(), outputs,
outputs.get_shape().as_list())
self.assertLess(error, 1e-4)