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_soft_copy_param_shape_check(self):
a = nn.Linear(2, 5)
b = nn.Linear(1, 5)
# Different shape
with self.assertRaises(AssertionError):
copy_param.soft_copy_param(a, b, 0.1)
with self.assertRaises(AssertionError):
copy_param.soft_copy_param(b, a, 0.1)
def test_soft_copy_param_scalar(self):
a = nn.Module()
a.p = nn.Parameter(torch.as_tensor(0.5))
b = nn.Module()
b.p = nn.Parameter(torch.as_tensor(1.0))
# a = (1 - tau) * a + tau * b
copy_param.soft_copy_param(target_link=a, source_link=b, tau=0.1)
torch_assert_allclose(a.p, torch.full_like(a.p, 0.55))
torch_assert_allclose(b.p, torch.full_like(b.p, 1.0))
copy_param.soft_copy_param(target_link=a, source_link=b, tau=0.1)
torch_assert_allclose(a.p, torch.full_like(a.p, 0.595))
torch_assert_allclose(b.p, torch.full_like(b.p, 1.0))
def test_soft_copy_param(self):
a = nn.Linear(1, 5)
b = nn.Linear(1, 5)
with torch.no_grad():
a.weight.fill_(0.5)
b.weight.fill_(1)
# a = (1 - tau) * a + tau * b
copy_param.soft_copy_param(target_link=a, source_link=b, tau=0.1)
torch_assert_allclose(a.weight, torch.full_like(a.weight, 0.55))
torch_assert_allclose(b.weight, torch.full_like(b.weight, 1.0))
copy_param.soft_copy_param(target_link=a, source_link=b, tau=0.1)
torch_assert_allclose(a.weight, torch.full_like(a.weight, 0.595))
torch_assert_allclose(b.weight, torch.full_like(b.weight, 1.0))