diff --git a/dnn/src/common/argmxx_helper.h b/dnn/src/common/argmxx_helper.h
index 87c69be3..a59e26fc 100644
--- a/dnn/src/common/argmxx_helper.h
+++ b/dnn/src/common/argmxx_helper.h
@@ -56,7 +56,7 @@ struct ArgmxxOp {
     ArgmxxOp(stype_ *src, dt_int32 *dst, uint32_t A, uint32_t B, uint32_t C):
         src(src), dst(dst), A(A), B(B), C(C),
         INIT(wtype(is_max ? DTypeTrait<stype_>::min() :
-                    DTypeTrait<stype_>::max(), -1))
+                    DTypeTrait<stype_>::max(), 0))
     {
     }
     MEGDNN_HOST MEGDNN_DEVICE wtype read(uint32_t idx)
diff --git a/dnn/src/naive/argmxx/opr_impl.cpp b/dnn/src/naive/argmxx/opr_impl.cpp
index 5754303a..8cca17ae 100644
--- a/dnn/src/naive/argmxx/opr_impl.cpp
+++ b/dnn/src/naive/argmxx/opr_impl.cpp
@@ -45,7 +45,7 @@ void exec_forward(_megdnn_tensor_in src,
     reduce::get_ABC(src.layout, A, B, C, param.axis);
     for (size_t a = 0; a < A; ++a) for (size_t c = 0; c < C; ++c) {
         float best_val = traits<is_max>::init;
-        size_t best_arg = -1;
+        size_t best_arg = 0;
         for (size_t b = 0; b < B; ++b) {
             float curr_val = float(src.ptr<T>()[(a*B+b)*C+c]);
             if (traits<is_max>::better_than(curr_val, best_val)) {
diff --git a/imperative/python/test/unit/functional/test_functional.py b/imperative/python/test/unit/functional/test_functional.py
index 76ca492d..acdf1a51 100644
--- a/imperative/python/test/unit/functional/test_functional.py
+++ b/imperative/python/test/unit/functional/test_functional.py
@@ -527,3 +527,20 @@ def test_nms_is_same():
     assert op3 != op4
 
 
+
+
+def test_argmxx_on_inf():
+    def run_argmax():
+        x = F.zeros((100, 100))
+        x[:] = -float("inf")
+        idxs = F.argmax(x, axis=0)
+        return idxs
+
+    def run_argmin():
+        x = F.zeros((100, 100))
+        x[:] = float("inf")
+        idxs = F.argmin(x, axis=0)
+        return idxs
+
+    assert all(run_argmax() >= 0)
+    assert all(run_argmin() >= 0)