def test_chainer_parameter_grad_setter(shape):
arr = numpy.full(shape, 17, 'float32')
chainer_param = chainer.Parameter(arr)
# Conversion
torch_param = cpm.ChainerParameter(chainer_param)
# Initialize grad
torch_param.requires_grad = True
optimizer = torch.optim.SGD([torch_param], lr=0.01, momentum=0.9)
optimizer.zero_grad()
# Setter
grad = torch.full(shape, 9, dtype=torch.float32)
torch_param.grad = grad
numpy.testing.assert_array_equal(grad, torch_param.grad)
def test_chainer_parameter(shape):
# initialized parameter
arr = numpy.full(shape, 17, 'float32')
chainer_param = chainer.Parameter(arr)
# Conversion
torch_param = cpm.ChainerParameter(chainer_param)
assert isinstance(torch_param, torch.nn.Parameter)
assert torch_param.shape == shape
assert (torch_param.data.numpy() == numpy.full(shape, 17, 'float32')).all()
# Test memory sharing
new_arr = numpy.random.randint(-4, 4, shape)
torch_param.data[...] = torch.tensor(new_arr.copy())
assert (chainer_param.array == new_arr).all()
def test_chainer_parameter_grad_getter(shape):
arr = numpy.full(shape, 17, 'float32')
grad = numpy.full(shape, 9, 'float32')
chainer_param = chainer.Parameter(arr)
chainer_param.grad = grad.copy()
# Conversion
torch_param = cpm.ChainerParameter(chainer_param)
# Getter
torch_grad = torch_param.grad
assert isinstance(torch_grad, torch.Tensor)
assert (torch_grad.numpy() == grad).all()
# Test memory sharing
new_arr = numpy.random.randint(-4, 4, shape)
torch_grad[...] = torch.tensor(new_arr.copy())
assert (chainer_param.grad == new_arr).all()
def test_chainer_parameter_uninitialized():
# Uninitialized parameters are not supported
chainer_param = chainer.Parameter()
with pytest.raises(TypeError):
cpm.ChainerParameter(chainer_param)