use shared_ptr to capture array in lambda

cpp/pybind/core/tensor_converter.cpp

@@ -106,10 +106,10 @@ Tensor PyArrayToTensor(py::array array, bool inplace) {
     Dtype dtype = pybind_utils::ArrayFormatToDtype(info.format, info.itemsize);
     Device device("CPU:0");
 
-    array.inc_ref();
-    std::function<void(void*)> deleter = [array](void*) -> void {
+    auto shared_array = std::make_shared<py::array>(array);
+    std::function<void(void*)> deleter = [shared_array](void*) mutable -> void {
         py::gil_scoped_acquire acquire;
-        array.dec_ref();
+        shared_array.reset();
     };
     auto blob = std::make_shared<Blob>(device, info.ptr, deleter);
     Tensor t_inplace(shape, strides, info.ptr, dtype, blob);