diff --git a/imperative/python/megengine/core/tensor/indexing.py b/imperative/python/megengine/core/tensor/indexing.py
index f55136392c31e9a2a545d9632d92624c6b4323a0..4dc1da3bdb3e6eeced6424b11ad2919b5195922b 100644
--- a/imperative/python/megengine/core/tensor/indexing.py
+++ b/imperative/python/megengine/core/tensor/indexing.py
@@ -119,12 +119,16 @@ def unpack_getitem(inp, tuple_val, *, allow_newaxis=True):
                 else 1
             )
     else:
-        if ndim_indexed > inp.ndim:
-            raise IndexError(
-                "too many indices for tensor: tensor is {}-dimensional, but {} were indexed".format(
-                    inp.ndim, len(tuple_val)
+        try:
+            if ndim_indexed > inp.ndim:
+                raise IndexError(
+                    "too many indices for tensor: tensor is {}-dimensional, but {} were indexed".format(
+                        inp.ndim, len(tuple_val)
+                    )
                 )
-            )
+        except ValueError:
+            # ignore
+            pass
 
     tuple_val = remove_ellipsis(inp, tuple_val)
     use_subtensor = True
diff --git a/imperative/python/src/tensor.cpp b/imperative/python/src/tensor.cpp
index 8cede64a80915bc07efa23fa5b01792cf1d7aa4c..d81de041503bec44b75be3a70fdfe39f69cf8df5 100644
--- a/imperative/python/src/tensor.cpp
+++ b/imperative/python/src/tensor.cpp
@@ -272,16 +272,12 @@ PyObject* TensorWrapper::device() {
 
 PyObject* TensorWrapper::numpy() {
     auto hv = m_tensor->numpy();
-    // if (!hv) {
-    //     PyErr_SetString(PyExc_ValueError, "tensor invalid");
-    //     return nullptr;
-    // }
-    auto arr = py::reinterpret_steal<py::array>(
-            npy::ndarray_from_tensor(hv->as_nd(true), npy::ShareType::TRY_SHARE));
-    if (!arr) {
+    if (!hv) {
         PyErr_SetString(PyExc_ValueError, "tensor invalid");
         return nullptr;
     }
+    auto arr = py::reinterpret_steal<py::array>(
+            npy::ndarray_from_tensor(hv->as_nd(true), npy::ShareType::TRY_SHARE));
     if (hv->shape().is_scalar()) {
         mgb_assert(PyArray_Check(arr.ptr()));
         return PyArray_Squeeze(reinterpret_cast<PyArrayObject*>(arr.ptr()));
diff --git a/imperative/src/impl/transformations/scalar.cpp b/imperative/src/impl/transformations/scalar.cpp
index 891abdd7c319dcbb9cd6ec1c907398e80cf71dc5..8daa5827ea47d6fb9093adf7a2098003477c64bd 100644
--- a/imperative/src/impl/transformations/scalar.cpp
+++ b/imperative/src/impl/transformations/scalar.cpp
@@ -51,6 +51,7 @@ bool is_scalar_shape(ValueRef shape) {
     if (shape.is<ScalarValue>()) {
         return false;
     }
+    // may have performance issue
     auto shape_of_shape = shape.shape();
     if (!shape_of_shape) {
         // assume not scalar
@@ -211,14 +212,21 @@ std::vector<ValueRef> subtensor_rule(
         const Subtensor& subtensor, Span<ValueRef> inputs) {
     mgb_assert(inputs.size() >= 1);
     auto input = inputs[0];
-    size_t ndim = input.is<ScalarValue>() ? 0 : input.shape()->ndim;
-    for (auto&& [axis, begin, end, step, idx] : subtensor.items) {
-        if (idx) {
-            ndim--;
+    bool is_scalar;
+    mgb_assert(!input.is<ScalarValue>(), "subtensor shouldn't have scalar input");
+    if (auto shape = input.shape()) {
+        size_t ndim = input.shape()->ndim;
+        for (auto&& [axis, begin, end, step, idx] : subtensor.items) {
+            if (idx) {
+                ndim--;
+            }
         }
+        is_scalar = ndim == 0;
+    } else {
+        is_scalar = false;
     }
     auto output = imperative::apply(subtensor, unwrap_inputs(inputs))[0];
-    if (!ndim) {
+    if (is_scalar) {
         return {ScalarValue::make(output)};
     } else {
         return {output};
@@ -261,8 +269,7 @@ std::vector<ValueRef> fastpath_copy_rule(
 
 std::vector<ValueRef> reshape_rule(const Reshape& reshape, Span<ValueRef> inputs) {
     mgb_assert(inputs.size() == 2);
-    bool is_scalar =
-            (!inputs[1].is<ScalarValue>()) && *inputs[1].shape() == ValueShape{0};
+    bool is_scalar = is_scalar_shape(inputs[1]);
     auto unwrapped_input = inputs[0].is<ScalarValue>()
                                  ? inputs[0].cast<ScalarValue>().value()
                                  : inputs[0];