Fix error in Pad op convert (#3778)

* Fix bug in pad converter, value only needed in constant mode.
Signed-off-by: NYu Liu <yu.liu@enflame-tech.com>

* Fix bug in pad converter, value only needed in constant mode or default (constant).
Signed-off-by: NYu Liu <yu.liu@enflame-tech.com>

* Fix bug in pad converter. input_rank maybe 0 and will set wrong t_pads.sizes().
Signed-off-by: NYu Liu <yu.liu@enflame-tech.com>

* Fix bug in pad converter. input_rank maybe 0 and will set wrong t_pads.sizes().
Signed-off-by: NYu Liu <yu.liu@enflame-tech.com>
Co-authored-by: NYu Liu <yu.liu@enflame-tech.com>
Co-authored-by: NG. Ramalingam <grama@microsoft.com>

onnx/version_converter/adapters/pad_10_11.h

@@ -17,12 +17,11 @@ class Pad_10_11 final : public Adapter {
       // Turn pads attribute into input
       Tensor t_pads;
       t_pads.elem_type() = TensorProto_DataType_INT64;
-      int input_rank = node->inputs()[0]->sizes().size();
-      t_pads.sizes() = std::vector<int64_t> {2 * input_rank};
       auto& data_pads = t_pads.int64s();
       for (int64_t shape : node->is(kpads)) {
         data_pads.emplace_back(shape);
       }
+      t_pads.sizes() = std::vector<int64_t> {(int64_t)data_pads.size()};
       Value* v_pads = graph->addInitializerAndInput(t_pads);
       node->addInput(v_pads);
       node->removeAttribute(kpads);