Limit outputs for accuracy checker

Always match dims with the smaller dimension from onnxruntime. Retina net
is outputting the correct amount of data (300) but out of onnxruntime, the
output is much smaller (5).

src/targets/gpu/lowering.cpp

@@ -111,6 +111,7 @@ struct miopen_apply
         add_loop_op();
         add_neg_op();
         add_nms_op();
+        //add_nonzero_op();
         add_select_module_op();
     }
 
@@ -360,6 +361,26 @@ struct miopen_apply
         });
     }
 
+    void add_nonzero_op()
+    {
+        apply_map.emplace("nonzero", [=](instruction_ref ins) {
+            auto s      = ins->get_shape();
+            auto output = insert_allocation(ins, s);
+            std::vector<instruction_ref> cpu_inputs;
+            auto inputs = ins->inputs();
+            std::transform(
+                inputs.begin(), inputs.end(), std::back_inserter(cpu_inputs), [&](auto in) {
+                    return mod->insert_instruction(ins, make_op("hip::copy_from_gpu"), in);
+                });
+            cpu_inputs.front() =
+                mod->insert_instruction(ins, make_op("hip::sync_stream"), cpu_inputs);
+            auto cpu_out = mod->insert_instruction(ins, ins->get_operator(), cpu_inputs);
+            auto gpu_out =
+                mod->insert_instruction(ins, make_op("hip::copy_to_gpu"), cpu_out, output);
+            return mod->replace_instruction(ins, gpu_out);
+        });
+    }
+
     /**
      * Adds dynamic allocation for submodule output parameter.
      */

tools/accuracy/accuracy_checker.py

@@ -93,7 +93,7 @@ def check_correctness(gold_outputs,
                       rtol=1e-3,
                       atol=1e-3,
                       verbose=False):
-    if len(gold_outputs) != len(outputs):
+    if len(gold_outputs) > len(outputs):
         print('Number of outputs {} is not equal to expected number {}'.format(
             len(outputs), len(gold_outputs)))
         return False