merge develop

doc/fluid/dev/use_eigen_cn.md

@@ -7,7 +7,7 @@
 
 Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。
 
-关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
+关于Eigen Tensor模块的详细介绍请参考[Eigen文档](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
 
 
 ## paddle::framework::Tensor

paddle/fluid/operators/sum_mkldnn_op.cc

@@ -34,15 +34,15 @@
 namespace paddle {
 namespace operators {
 
-using paddle::framework::Tensor;
-using paddle::platform::MKLDNNDeviceContext;
-using paddle::platform::CPUDeviceContext;
 using framework::DataLayout;
 using mkldnn::memory;
 using mkldnn::primitive;
+using mkldnn::reorder;
 using mkldnn::stream;
 using mkldnn::sum;
-using mkldnn::reorder;
+using paddle::framework::Tensor;
+using paddle::platform::CPUDeviceContext;
+using paddle::platform::MKLDNNDeviceContext;
 using platform::to_void_cast;
 
 template <typename T>
@@ -175,18 +175,35 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         auto& sel_row = get_selected_row(i);
         first_dim += sel_row.rows().size();
       }
-      auto in_dim =
-          framework::vectorize(get_selected_row(N - 1).value().dims());
+
+      std::vector<int64_t> in_dim;
+      for (int i = 0; i < N; i++) {
+        auto& sel_row = get_selected_row(i);
+        if (sel_row.rows().size() > 0) {
+          in_dim = framework::vectorize(sel_row.value().dims());
+          break;
+        }
+      }
+
+      if (in_dim.empty()) {
+        VLOG(3) << "WARNING: all the inputs are empty";
+        in_dim = framework::vectorize(get_selected_row(N - 1).value().dims());
+      } else {
+        in_dim[0] = static_cast<int64_t>(first_dim);
+      }
+
       in_dim[0] = static_cast<int64_t>(first_dim);
 
       out_value->Resize(framework::make_ddim(in_dim));
 
+      out_value->mutable_data<T>(ctx.GetPlace());
+
       // if all the input sparse vars are empty, no need to
       // merge these vars.
       if (first_dim == 0UL) {
         return;
       }
-      out_value->mutable_data<T>(ctx.GetPlace());
+
       math::SelectedRowsAddTo<CPUDeviceContext, T> functor;
       int64_t offset = 0;
       for (int i = 0; i < N; i++) {

paddle/fluid/operators/sum_op.h

@@ -105,18 +105,30 @@ class SumKernel : public framework::OpKernel<T> {
         auto &sel_row = get_selected_row(i);
         first_dim += sel_row.rows().size();
       }
-      auto in_dim =
-          framework::vectorize(get_selected_row(N - 1).value().dims());
+
+      std::vector<int64_t> in_dim;
+      for (int i = 0; i < N; i++) {
+        auto &sel_row = get_selected_row(i);
+        if (sel_row.rows().size() > 0) {
+          in_dim = framework::vectorize(sel_row.value().dims());
+          break;
+        }
+      }
+      if (in_dim.empty()) {
+        VLOG(3) << "WARNING: all the inputs are empty";
+        in_dim = framework::vectorize(get_selected_row(N - 1).value().dims());
+      } else {
         in_dim[0] = static_cast<int64_t>(first_dim);
+      }
 
       out_value->Resize(framework::make_ddim(in_dim));
+      out_value->mutable_data<T>(context.GetPlace());
 
       // if all the input sparse vars are empty, no need to
       // merge these vars.
       if (first_dim == 0UL) {
         return;
       }
-      out_value->mutable_data<T>(context.GetPlace());
 
       math::SelectedRowsAddTo<DeviceContext, T> functor;
 

python/paddle/fluid/tests/unittests/test_sum_op.py

@@ -17,6 +17,8 @@ from __future__ import print_function
 import unittest
 import numpy as np
 from op_test import OpTest
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
 
 
 class TestSumOp(OpTest):
@@ -42,5 +44,66 @@ class TestSumOp(OpTest):
         pass
 
 
+class TestSelectedRowsSumOp(OpTest):
+    def check_with_place(self, place):
+        scope = core.Scope()
+        self.check_input_and_optput(scope, place, True, True, True)
+        self.check_input_and_optput(scope, place, False, True, True)
+        self.check_input_and_optput(scope, place, False, False, True)
+        self.check_input_and_optput(scope, place, False, False, False)
+
+    def check_input_and_optput(self,
+                               scope,
+                               place,
+                               w1_has_data=False,
+                               w2_has_data=False,
+                               w3_has_data=False):
+
+        self.create_selected_rows(scope, place, "W1", w1_has_data)
+        self.create_selected_rows(scope, place, "W2", w2_has_data)
+        self.create_selected_rows(scope, place, "W3", w3_has_data)
+
+        # create Out Variable
+        out = scope.var('Out').get_selected_rows()
+
+        # create and run sum operator
+        sum_op = Operator("sum", X=["W1", "W2", "W3"], Out='Out')
+        sum_op.run(scope, place)
+
+        has_data_w_num = 0
+        for w in [w1_has_data, w2_has_data, w3_has_data]:
+            if not w:
+                has_data_w_num += 1
+
+        self.assertEqual(7 * has_data_w_num, len(out.rows()))
+
+    def create_selected_rows(self, scope, place, var_name, isEmpty):
+        # create and initialize W Variable
+        if not isEmpty:
+            rows = [0, 1, 2, 3, 4, 5, 6]
+            row_numel = 12
+        else:
+            rows = []
+            row_numel = 12
+
+        var = scope.var(var_name)
+        w_selected_rows = var.get_selected_rows()
+        w_selected_rows.set_height(len(rows))
+        w_selected_rows.set_rows(rows)
+        w_array = np.ones((len(rows), row_numel)).astype("float32")
+        for i in range(len(rows)):
+            w_array[i] *= i
+        w_tensor = w_selected_rows.get_tensor()
+        w_tensor.set(w_array, place)
+
+        return var
+
+    def test_w_is_selected_rows(self):
+        places = [core.CPUPlace()]
+        # currently only support CPU
+        for place in places:
+            self.check_with_place(place)
+
+
 if __name__ == "__main__":
     unittest.main()