def sync_parameters(self):
copy_param.copy_param(target_link=self.model, source_link=self.shared_model)
copy_param.soft_copy_param(
target_link=self.shared_average_model,
source_link=self.model,
tau=1 - self.trust_region_alpha,
)
def test_copy_param_shape_check(self):
a = nn.Linear(2, 5)
b = nn.Linear(1, 5)
with self.assertRaises(RuntimeError):
# Different shape
copy_param.copy_param(a, b)
with self.assertRaises(RuntimeError):
# Different shape
copy_param.copy_param(b, a)
def test_copy_param_scalar(self):
a = nn.Module()
a.p = nn.Parameter(torch.Tensor([1]))
b = nn.Module()
b.p = nn.Parameter(torch.Tensor([2]))
self.assertNotEqual(a.p.detach().numpy(), b.p.detach().numpy())
# Copy b's parameters to a
copy_param.copy_param(a, b)
self.assertEqual(a.p.detach().numpy(), b.p.detach().numpy())
def test_copy_param(self):
a = nn.Linear(1, 5)
b = nn.Linear(1, 5)
s = torch.from_numpy(np.random.rand(1, 1).astype(np.float32))
a_out = list(a(s).detach().numpy().ravel())
b_out = list(b(s).detach().numpy().ravel())
self.assertNotEqual(a_out, b_out)
# Copy b's parameters to a
copy_param.copy_param(a, b)
a_out_new = list(a(s).detach().numpy().ravel())
b_out_new = list(b(s).detach().numpy().ravel())
self.assertEqual(a_out_new, b_out)
self.assertEqual(b_out_new, b_out)
def load(self, dirname):
super().load(dirname)
copy_param.copy_param(target_link=self.shared_model,
source_link=self.model)
def sync_parameters(self):
copy_param.copy_param(target_link=self.model,
source_link=self.shared_model)