def check_double_backward(self, m_data, v_data, y_grad, m_grad_grad,
v_grad_grad):
gaussian = functions.Gaussian()
def f(m, v):
return gaussian.apply((m, v))
gradient_check.check_double_backward(
f, (m_data, v_data), y_grad, (m_grad_grad, v_grad_grad),
atol=5e-4, rtol=5e-3)
def check_backward(self, m_data, v_data, y_grad):
# Instantiate the FunctionNode object outside the function that is
# tested in order to reuse the same noise for the numerical gradient
# computations (the noise is generated once during its first forward
# pass, then reused)
# TODO(hvy): Do no expose interals of the tested function using
# e.g. numpy.random.RandomState
gaussian = functions.Gaussian()
def f(m, v):
return gaussian.apply((m, v))[0]
gradient_check.check_backward(f, (m_data, v_data), y_grad,
**self.check_backward_options)
def check_backward(self, m_data, v_data, y_grad):
gradient_check.check_backward(functions.Gaussian(), (m_data, v_data),
y_grad, **self.check_backward_options)
def check_backward(self, m_data, v_data, y_grad):
gradient_check.check_backward(
functions.Gaussian(), (m_data, v_data), y_grad,
atol=1e-4, rtol=1e-3)