def test_tensor_used_in_backward(self):
# if tensor is used in backward, its ref_count should increase 1
a = flow.Tensor(4.1, require_grad=True)
b = flow.Tensor(3.1, require_grad=True)
d = a * b
# one for mul operation and one for backward of d
assert a.forward_ref_count == 1
assert a.backward_ref_count == 1
def test_iadd(self):
a = flow.Tensor(3)
b = a
a += flow.Tensor(5)
assert np.allclose(a.data, b.data)
a = flow.Tensor(3)
b = a
a = a + flow.Tensor(5)
assert not np.allclose(a.data, b.data)
def test_imul(self):
a = flow.Tensor(3)
b = a
a *= flow.Tensor(5)
assert np.allclose(a.data, b.data)
a = flow.Tensor(3)
b = a
a = a * flow.Tensor(5)
assert not np.allclose(a.data, b.data)
def test_isub(self):
a = flow.Tensor(3)
b = a
a -= flow.Tensor(5)
assert np.allclose(a.data, b.data)
a = flow.Tensor(3)
b = a
a = a - flow.Tensor(5)
assert not np.allclose(a.data, b.data)
def test_itrue_div(self):
# inplace div will cause a dtype change from int to float, which raise error
# therefore define a to be 3.1 in the first place
a = flow.Tensor(3.1)
b = a
a /= flow.Tensor(5)
assert np.allclose(a.data, b.data)
a = flow.Tensor(3)
b = a
a = a / flow.Tensor(5)
assert not np.allclose(a.data, b.data)
def test_function_inside_function(self):
# if a autograd function A is called inside of a autograd function B
# autograd function A should no increase ref_count
class trivial_function(flow.autograd.Function):
@staticmethod
def forward(ctx, tensor):
return F.add(tensor, flow.Tensor(0))
@staticmethod
def backward(ctx, grad):
return F.add(grad, flow.Tensor(0))
a = flow.Tensor(4.1, require_grad=True)
b = trivial_function.apply(a)
assert a.forward_ref_count == 1
def __init__(self):
super().__init__()
self.a = Identity(flow.Tensor([[4.0, 5.0]], require_grad=True))
self.b = Identity(flow.Tensor([[5.0], [6.0]], require_grad=True))
self.c = Identity(
flow.Tensor([[1.0, 2.0], [3.0, 4.0]], require_grad=True))
def test_ref_count(self):
a = flow.Tensor(4.1, require_grad=True)
b = a + 1
c = a + 2
d = a + 3
assert a.forward_ref_count == 3
def backward(ctx, grad):
return F.add(grad, flow.Tensor(0))
def forward(ctx, tensor):
return F.add(tensor, flow.Tensor(0))