pass sgd at first iter

paddle/operators/parallel_do_op.cc

@@ -185,18 +185,27 @@ class ParallelDoGradOp : public OperatorBase {
     // merge grad
     for (auto &s : Outputs(framework::GradVarName(kParameters))) {
       LOG(INFO) << s;
-      // std::string s_buf = s + "@BUF";
+
-      // auto *t_buf = sub_scopes[0]->Var(s_buf)->GetMutable<LoDTensor>();
+      auto &t = sub_scopes[0]->FindVar(s)->Get<LoDTensor>();
+      LOG(INFO) << t;
+
+      std::string s_buf = s + "@BUF";
+      auto *t_buf = sub_scopes[0]->Var(s_buf)->GetMutable<LoDTensor>();
+
       for (size_t place_idx = 1; place_idx < places.size(); ++place_idx) {
+        auto &tt = sub_scopes[place_idx]->FindVar(s)->Get<LoDTensor>();
         LOG(INFO) << place_idx;
-        LOG(INFO) << sub_scopes[place_idx]->FindVar(s)->Get<LoDTensor>();
+        LOG(INFO) << tt;
-        // Copy grad[i] to grad_buf[0]
+        framework::CopyFrom(tt, places[0], t_buf);
 
-        // sum_op
+        auto sum_op = framework::OpRegistry::CreateOp(
+            "sum", {{"X", {s, s_buf}}}, {{"Out", {s}}},
+            framework::AttributeMap{});
+        sum_op->Run(*sub_scopes[0], place);
       }
 
-      // Copy grad[0] to grad
+      LOG(INFO) << t;
-      // auto *t = scope.FindVar(s)->GetMutable<LoDTensor>();
+      framework::CopyFrom(t, place, scope.FindVar(s)->GetMutable<LoDTensor>());
     }
   }
 };

python/paddle/v2/fluid/tests/test_parallel_op.py

@@ -27,7 +27,8 @@ class ParallelOpTest(unittest.TestCase):
             pd.write_output(hidden)
         data = pd()
         loss = layers.mean(x=data)
-        append_backward(loss)
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+        sgd_optimizer.minimize(loss)
 
         exe = fluid.Executor(fluid.CPUPlace())
         exe.run(fluid.default_startup_program())