fixed bugs in gru implementation.

src/onnx/onnx.cpp

@@ -734,50 +734,53 @@ struct onnx_parser
             dirct = op::gru::reverse;
         }
 
-        std::vector<std::string> actv_func_names = {"sigmoid", "tanh"};
+        std::vector<std::string> vec_names = {"sigmoid", "tanh"};
         if(contains(attributes, "activations"))
         {
             auto names = attributes.at("activations").strings();
-            actv_func_names.clear();
+            vec_names.clear();
             for(auto& fn : names)
             {
-                actv_func_names.push_back(fn);
+                vec_names.push_back(fn);
             }
         }
 
-        if(actv_func_names.size() != 2)
-        {
-            MIGRAPHX_THROW("GRU: wrong activation function attribute");
-        }
-
         // need 4 activation functions
         if(dirct == op::gru::bidirectional)
         {
-            // one name is provided, need to repeat the function 3 times
-            if(actv_func_names.size() == 1)
+            // 4 activation functions are used in the bidirectional 
+            // scenario. No spec is provided in onnx::operator. we
+            // use the algorithm that: if 1 actv function is provides,
+            // repeat 1 four times. If 2 actv functins are provides,
+            // assume forward and reverse use the same pair of actv
+            // functions. For the case of 3 actv functions provided,
+            // assume the 3rd one is repeated once and used by the 
+            // reverse direction. 
+            // This may need change later 
+            if(vec_names.size() == 1)
             {
-                actv_func_names.resize(4, actv_func_names.at(0));
+                vec_names.insert(vec_names.end(), 3, vec_names.at(0));
             }
-            else if(actv_func_names.size() == 2)
+            else if(vec_names.size() == 2)
             {
-                actv_func_names.insert(
-                    actv_func_names.end(), actv_func_names.begin(), actv_func_names.end());
+                // repeat the activation functions
+                vec_names.push_back(vec_names.at(0));
+                vec_names.push_back(vec_names.at(1));
             }
-            else if(actv_func_names.size() == 3)
+            else if(vec_names.size() == 3)
             {
-                MIGRAPHX_THROW(
-                    "GRU: birectional network cannot have 3 activation functions in attribute");
+                vec_names.push_back(vec_names.at(2));
             }
         }
         else
         {
-            if(actv_func_names.size() == 1)
+            if(vec_names.size() == 1)
             {
-                actv_func_names.push_back(actv_func_names.at(0));
+                vec_names.push_back(vec_names.at(0));
             }
         }
 
-        for_each(actv_func_names.begin(), actv_func_names.end(), [&](auto& name) {
+        for_each(vec_names.begin(), vec_names.end(), [&](auto& name) {
             if(map_actv_funcs.count(name) == 0)
             {
                 MIGRAPHX_THROW("GRU: activation function " + name + " not supported");
@@ -785,7 +788,7 @@ struct onnx_parser
         });
 
         std::vector<operation> vec_actv_funcs;
-        for_each(actv_func_names.begin(), actv_func_names.end(), [&](auto& name) {
+        for_each(vec_names.begin(), vec_names.end(), [&](auto& name) {
             vec_actv_funcs.push_back(map_actv_funcs[name]);
         });
 

src/rewrite_gru.cpp

@@ -103,17 +103,15 @@ void rewrite_gru::apply(program& prog) const
                                         gru_op.actv_funcs.at(2),
                                         gru_op.actv_funcs.at(3));
 
-            auto final_output =
-                prog.insert_instruction(ins, op::concat{0}, ret_forward[1], ret_reverse[1]);
+            //auto final_output =
+            //    prog.insert_instruction(ins, op::concat{0}, ret_forward[1], ret_reverse[1]);
 
             // add the dimension of num_direction
             ret_forward[0] = prog.insert_instruction(ins, op::unsqueeze{{1}}, ret_forward[0]);
             ret_reverse[0] = prog.insert_instruction(ins, op::unsqueeze{{1}}, ret_reverse[0]);
 
             // concat the forward and reverse output
-            auto replaced_arg =
-                prog.replace_instruction(ins, op::concat{1}, {ret_forward[0], ret_reverse[0]});
-            replaced_arg->add_output(final_output);
+            prog.replace_instruction(ins, op::concat{1}, {ret_forward[0], ret_reverse[0]});
         }
         else
         {
@@ -153,8 +151,7 @@ void rewrite_gru::apply(program& prog) const
                                 gru_op.actv_funcs.at(1));
 
             // add the dimension of num_direction
-            auto replaced_arg = prog.replace_instruction(ins, op::unsqueeze{{1}}, ret[0]);
-            replaced_arg->add_output(ret[1]);
+            prog.replace_instruction(ins, op::unsqueeze{{1}}, ret[0]);
         }
     }
 }