nikita-savelyevv / ennf Goto Github PK

in module nn_lib\tests\test_tensor_backward.py:
function
def test_exponent_gradient_scalar_2(self):
a = Tensor(3.1, requires_grad=True)
b = F.exp(a)
b.backward(Tensor(4))
self.assertAlmostEqual(a.grad.data, 4*np.exp(3.1), 5)
self.assertEqual(b.grad.data, 4)
in first assert give "not equal" output:

raise self.failureException(msg)
AssertionError: array(88.79179, dtype=float32) != 88.79180512576654 within 5 places (1.893191888768797e-05 difference)

i think issue with float types of default python float values and numpy float32. Default float type is float == float64 in numpy. But when we create Tensor instance, we use float32 type of numpy.

if we use float64 type in assertion, when we use float32 in Tensor not correct.

My solve this problem is change
row:
self.assertAlmostEqual(a.grad.data, 4*np.exp(3.1), 5)
to
self.assertAlmostEqual(a.grad.data, np.dot(4, np.exp(np.array(3.1, np.float32))), 5)
and test complete successfully!