update

paddle/fluid/framework/selected_rows.cc

@@ -38,10 +38,10 @@ struct ReAllocateVisitor {
   framework::DDim dims_;
 };
 
-struct TensorSlicedCopyVisitor {
-  TensorSlicedCopyVisitor(const platform::Place& place, framework::Tensor* dst,
-                          int64_t dst_offset, const framework::Tensor src,
-                          int64_t src_offset, int64_t size)
+struct TensorCopyVisitor {
+  TensorCopyVisitor(const platform::Place& place, framework::Tensor* dst,
+                    int64_t dst_offset, const framework::Tensor src,
+                    int64_t src_offset, int64_t size)
       : place_(place),
         dst_(dst),
         dst_offset_(dst_offset),
@@ -121,10 +121,27 @@ bool SelectedRows::HasKey(int64_t key) const {
                                                                    : true;
 }
 
-Tensor SelectedRows::Get(int64_t key) const {
+bool SelectedRows::Get(int64_t key, framework::Tensor* value,
+                       int64_t row) const {
   int64_t index = Index(key);
   PADDLE_ENFORCE_GE(index, 0, "The key should be exists in the Table.");
-  return value_->Slice(index, index + 1);
+  PADDLE_ENFORCE(value->IsInitialized(),
+                 "The value tensor should be initialized.");
+
+  int64_t value_width = value->numel() / value->dims()[0];
+  PADDLE_ENFORCE_EQ(value_width, value_->numel() / value_->dims()[0],
+                    "output tensor should have the same shape with table "
+                    "execpt the dims[0].");
+
+  // TODO(Yancey1989): support other place
+  platform::CPUPlace cpu;
+
+  framework::VisitDataType(
+      framework::ToDataType(value_->type()),
+      TensorCopyVisitor(cpu, value, row * value_width, *value_.get(),
+                        index * value_width, value_width));
+
+  return true;
 }
 
 bool SelectedRows::Set(int64_t key, const framework::Tensor& value) {
@@ -143,7 +160,7 @@ bool SelectedRows::Set(int64_t key, const framework::Tensor& value) {
     rows_.push_back(key);
     index = rows_.size() - 1;
     is_new_key = true;
-    // whether need to resize the value
+    // whether need to resize the table
     if (static_cast<int64_t>(rows_.size()) > value_->dims()[0]) {
       auto dims = value_->dims();
       dims[0] = (dims[0] + 1) << 1;
@@ -154,9 +171,9 @@ bool SelectedRows::Set(int64_t key, const framework::Tensor& value) {
 
   framework::VisitDataType(
       framework::ToDataType(value.type()),
-      TensorSlicedCopyVisitor(cpu, value_.get(),
-                              index * value_->numel() / value_->dims()[0],
-                              value, static_cast<int64_t>(0), value.numel()));
+      TensorCopyVisitor(cpu, value_.get(),
+                        index * value_->numel() / value_->dims()[0], value,
+                        static_cast<int64_t>(0), value.numel()));
   return is_new_key;
 }
 

paddle/fluid/framework/selected_rows.h

@@ -62,9 +62,10 @@ class SelectedRows {
    * @brief Get a value by the specified key, if the
    * key does not exists, this function would throw an exception.
    *
-   * @return a sliced tensor
+   * @return true if the Get operation successed.
    */
-  Tensor Get(int64_t key) const;
+
+  bool Get(int64_t key, framework::Tensor* tensor, int64_t row = 0) const;
 
   /*
    * @brief Set a key-value pair into the table.

paddle/fluid/framework/selected_rows_test.cc

@@ -62,6 +62,10 @@ TEST_F(SelectedRowsTester, SerializeAndDeseralize) {
 TEST_F(SelectedRowsTester, Table) {
   platform::CPUPlace cpu;
   SelectedRows table;
+  // initialize a sparse table
+  table.mutable_value()->Resize(framework::make_ddim({1, 100}));
+  table.mutable_value()->mutable_data<float>(cpu);
+  table.mutable_rows()->push_back(1);
 
   int64_t key = 10000;
   framework::Tensor value;
@@ -69,15 +73,21 @@ TEST_F(SelectedRowsTester, Table) {
   auto ptr = value.mutable_data<float>(cpu);
   ptr[0] = static_cast<float>(10);
 
-  ASSERT_EQ(table.rows().size(), static_cast<size_t>(0));
+  ASSERT_EQ(table.rows().size(), static_cast<size_t>(1));
   ASSERT_EQ(table.HasKey(key), false);
 
   table.Set(key, value);
 
-  ASSERT_EQ(table.rows().size(), static_cast<size_t>(1));
+  ASSERT_EQ(table.rows().size(), static_cast<size_t>(2));
   ASSERT_EQ(table.HasKey(key), true);
-  ASSERT_EQ(table.value().dims()[0], static_cast<int64_t>(2));
-  ASSERT_EQ(table.Get(key).data<float>()[0], static_cast<float>(10));
+  // check re-allocate
+  ASSERT_EQ(table.value().dims()[0], static_cast<int64_t>(4));
+
+  framework::Tensor get_value;
+  get_value.mutable_data<float>(framework::make_ddim({20, 100}), cpu);
+  table.Get(key, &get_value, 10);
+
+  ASSERT_EQ(get_value.data<float>()[10 * 100], static_cast<float>(10));
 }
 
 }  // namespace framework

paddle/fluid/operators/sgd_op.h

@@ -107,7 +107,9 @@ class SGDOpKernel : public framework::OpKernel<T> {
       for (size_t i = 0; i < grad.rows().size(); i++) {
         PADDLE_ENFORCE(grad.rows()[i] < grad.height(),
                        "Input rows index should less than height");
-        int64_t id_index = param.index(grad.rows()[i]);
+        int64_t id_index = param.Index(grad.rows()[i]);
+        PADDLE_ENFORCE_GE(id_index, static_cast<int64_t>(0),
+                          "id should be in the table");
         for (size_t j = 0; j < grad_row_width; j++) {
           out_data[id_index * grad_row_width + j] -=
               lr[0] * grad_data[i * grad_row_width + j];