fix detach numpy in pytorch

Author: refraction-ray

Diff for: CHANGELOG.md

@@ -4,16 +4,20 @@
 
 ### Added
 
-- PyTorch backend support multi pytrees version of ``tree_map``
+- PyTorch backend support multi pytrees version of `tree_map`
 
 ### Changed
 
-- Refactor ``interfaces`` code as a submodule and add pytree support for args
+- Refactor `interfaces` code as a submodule and add pytree support for args
 
 - Change the way to register global setup internally, so that we can skip the list of all submodules
 
 - Refactor the tensortrans code to a pytree perspective
 
+### Fixed
+
+- Fixed `numpy` method bug in pytorch backend when the input tensor requires grad (#24)
+
 ## 0.2.1
 
 ### Added

Diff for: tensorcircuit/backends/pytorch_backend.py

@@ -285,6 +285,8 @@ def kron(self, a: Tensor, b: Tensor) -> Tensor:
     def numpy(self, a: Tensor) -> Tensor:
         if a.is_conj():
             return a.resolve_conj().numpy()
+        if a.requires_grad:
+            return a.detach().numpy()
         return a.numpy()
 
     def i(self, dtype: Any = None) -> Tensor:

Diff for: tests/test_interfaces.py

@@ -105,6 +105,32 @@ def f3(x):
     np.testing.assert_allclose(pg, 2 * np.ones([2]).astype(np.complex64), atol=1e-5)
 
 
+@pytest.mark.skipif(is_torch is False, reason="torch not installed")
+@pytest.mark.xfail(reason="see comment link below")
+@pytest.mark.parametrize("backend", [lf("tfb"), lf("jaxb")])
+def test_torch_interface_pytree(backend):
+    # pytree cannot support in pytorch autograd function...
+    # https://github.com/pytorch/pytorch/issues/55509
+    def f4(x):
+        return tc.backend.sum(x["a"] ** 2), tc.backend.sum(x["b"] ** 3)
+
+    f4_torch = tc.interfaces.torch_interface(f4, jit=False)
+    param4 = {
+        "a": torch.ones([2], requires_grad=True),
+        "b": torch.ones([2], requires_grad=True),
+    }
+
+    def f4_post(x):
+        r1, r2 = f4_torch(param4)
+        l4 = r1 + r2
+        return l4
+
+    pg = tc.get_backend("pytorch").grad(f4_post)(param4)
+    np.testing.assert_allclose(
+        pg["a"], 2 * np.ones([2]).astype(np.complex64), atol=1e-5
+    )
+
+
 @pytest.mark.parametrize("backend", [lf("npb"), lf("tfb"), lf("jaxb")])
 def test_scipy_interface(backend):
     n = 3