Enhance unit test.

Author: pkuyym

paddle/fluid/operators/sequence_expand_op.cc

@@ -33,10 +33,11 @@ class SequenceExpandOp : public framework::OperatorWithKernel {
                    "Output(Out) of SequenceExpandOp should not be null.");
 
     auto x_dims = ctx->GetInputDim("X");
+    auto out_dims = x_dims;
     int ref_level = ctx->Attrs().Get<int>("ref_level");
 
-    PADDLE_ENFORCE_EQ(x_dims.size(), 2U,
-                      "Dimension number of Input(X) should be 2.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Dimension number of Input(X) should be at least 2.");
 
     if (ctx->IsRuntime()) {
       framework::Variable* x_var =
@@ -50,15 +51,9 @@ class SequenceExpandOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE_LE(x_lod.size(), 1,
                         "Number of lod level of Input(X) should not be "
                         "greater than 1.");
-
-      PADDLE_ENFORCE(x_lod.size() == y_lod.size() || x_lod.size() == 0,
-                     "Level number of Input(X)'s lod should be either equal "
-                     "to 0 or equal to that of Input(Y).");
-
       PADDLE_ENFORCE_GT(y_lod.size(), 0,
                         "Level number of Input(Y)'s lod should be "
                         "greater than 0.");
-
       PADDLE_ENFORCE(
           ref_level == -1 ||
               (ref_level >= 0 && ref_level < static_cast<int>(y_lod.size())),
@@ -68,6 +63,14 @@ class SequenceExpandOp : public framework::OperatorWithKernel {
 
       if (ref_level == -1) ref_level = y_lod.size() - 1;
 
+      if (x_lod.size() > 0) {
+        PADDLE_ENFORCE(
+            x_lod.size() == 0 || x_lod[0].size() == y_lod[ref_level].size(),
+            "Level number of Input(X)'s lod should be 0. Otherwise "
+            "size of Input(X)'s first level lod should be equal to "
+            "size of Input(Y)'s lod of referred level.");
+      }
+
       int64_t out_first_dim = 0;
       if (y_lod[ref_level].size() <= 1) {
         out_first_dim = x_dims[0];
@@ -81,9 +84,12 @@ class SequenceExpandOp : public framework::OperatorWithKernel {
               (y_lod[ref_level][i] - y_lod[ref_level][i - 1]) * x_seq_len;
         }
       }
-      ctx->SetOutputDim("Out", {out_first_dim, x_dims[1]});
+      out_dims[0] = out_first_dim;
+      ctx->SetOutputDim("Out", out_dims);
     } else {
-      ctx->SetOutputDim("Out", {-1, x_dims[1]});
+      out_dims[0] = -1;
+      ctx->SetOutputDim("Out", out_dims);
+      ctx->ShareLoD("X", /*->*/ "Out");
     }
   }
 };
@@ -105,69 +111,69 @@ class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 Sequence Expand Operator.
 
-This operator expands input(X) according to LOD of input(Y).
+This operator expands `X` according to specified level lod of `Y`. Current
+implementation constaints that lod level of `X` should be at most 1. Attribute
+`ref_level` is used to specify which level lod of `Y` is referred to expand `X`.
+If set `ref_level` to -1, then last level lod of `Y` would be referred.
+Please note, rank of `X` should be at least 2, when the rank exceeds 2, `X`
+would be viewed as a 2-D tensor.
+
 Following are cases to better explain how this works:
+
 Case 1:
 
-Given a 2-level LoDTensor input(X)
-    X.lod = [[0,       2, 3],
-             [0, 1,    3, 4]]
-    X.data = [a, b, c, d]
+Given a 1-level LoDTensor input(X)
+    X.lod =  [[0,   2,        4]]
+    X.data = [[a], [b], [c], [d]]
     X.dims = [4, 1]
 and input(Y)
     Y.lod = [[0,    2,    4],
              [0, 3, 6, 7, 8]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 2-level LoDTensor
-    Out.lod = [[0,                2,    4],
-               [0,       3,       6, 7, 8]]
-    Out.data = [a, a, a, b, b, b, c, d]
+ref_level: 0
+then we get 1-level LoDTensor
+    Out.lod =  [[0,   2,        4,        6,        8]]
+    Out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
     Out.dims = [8, 1]
 
 Case 2:
 
+Given 1-level LoDTensor input(X)
+    X.lod =  [[0,   1,        4]]
+    X.data = [[a], [b], [c], [d]]
+    X.dims = [4, 1]
+and input(Y)
+    Y.lod = [[0,    2,    4],
+             [0, 3, 6, 6, 8]]
+ref_level: 0
+then we get 1-level LoDTensor
+    Out.lod =  [[0,   2,             5,             8]]
+    Out.data = [[a], [a], [b], [c], [d], [b], [c], [d]]
+    Out.dims = [8, 1]
+
+Case 3:
+
 Given a common Tensor input(X)
-    X.data = [a, b, c]
+    X.data = [[a], [b], [c]]
     X.dims = [3, 1]
 and input(Y)
     Y.lod = [[0, 2, 3, 6]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 1-level LoDTensor
-    Out.lod = [[0,    2, 3,      6]]
-    Out.data = [a, a, b, c, c, c]
+ref_level: -1
+then we a common Tensor
+    Out.data = [[a], [a], [b], [c], [c], [c]]
     Out.dims = [6, 1]
 
-Case 3:
+Case 4:
 
 Given a common Tensor input(X)
     X.data = [[a, b], [c, d], [e, f]]
     X.dims = [3, 2]
 and input(Y)
     Y.lod = [[0, 2, 3, 6]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 1-level LoDTensor
-    Out.lod = [[0,           2,     3,                     6]]
-    Out.data = [[a,b], [a,b] [c,d], [e, f], [e, f], [e, f]]
+ref_level: 0
+then we get a common LoDTensor
+    Out.data = [[a, b], [a, b] [c, d], [e, f], [e, f], [e, f]]
     Out.dims = [6, 2]
 
-Case 4:
-
-Given 2-level a LoDTensor input(X)
-    X.lod = [[0,       2, 3],
-             [0, 1,    3, 4]]
-    X.data = [a, b, c, d]
-    X.dims = [4, 1]
-and input(Y)
-    Y.lod = [[0,    2,    4],
-             [0, 3, 6, 6, 8]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 2-level LoDTensor
-    Out.lod = [[0,                2,    4],
-               [0,       3,       6, 6, 8]]
-    Out.data = [a, a, a, b, b, b, d, d]
-    Out.dims = [8, 1]
-
-
 )DOC");
   }
 };

paddle/fluid/operators/sequence_expand_op.h

@@ -22,6 +22,9 @@ namespace paddle {
 namespace operators {
 
 using LoDTensor = framework::LoDTensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename DeviceContext, typename T>
 class SequenceExpandKernel : public framework::OpKernel<T> {
@@ -30,15 +33,12 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
     auto* x = context.Input<LoDTensor>("X");
     auto* y = context.Input<LoDTensor>("Y");
     auto* out = context.Output<LoDTensor>("Out");
-    int ref_level = context.Attr<int>("ref_level");
 
-    out->mutable_data<T>(context.GetPlace());
+    int ref_level = context.Attr<int>("ref_level");
     auto& x_lod = x->lod();
     auto& y_lod = y->lod();
-
     PADDLE_ENFORCE_GT(y_lod.size(), 0,
                       "Level number of `Y`'s lod should be greater than 0.");
-
     PADDLE_ENFORCE(
         ref_level == -1 || (ref_level >= 0 && ref_level < y_lod.size()),
         "Invlid `ref_level`, which should be either equal to -1 "
@@ -47,6 +47,8 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
 
     if (ref_level == -1) ref_level = y_lod.size() - 1;
 
+    out->mutable_data<T>(context.GetPlace());
+
     if (y_lod[ref_level].size() <= 1) {
       framework::TensorCopy(*x, context.GetPlace(), out);
       return;
@@ -59,6 +61,8 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
     }
 
     int out_offset = 0;
+    auto& eigen_place =
+        *context.template device_context<DeviceContext>().eigen_device();
     for (size_t i = 1; i < y_lod[ref_level].size(); ++i) {
       int repeat_num = y_lod[ref_level][i] - y_lod[ref_level][i - 1];
       int x_start = i - 1;
@@ -68,16 +72,24 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
         x_end = x_lod[0][i];
       }
       int x_seq_len = x_end - x_start;
-      auto x_sub_tensor = x->Slice(x_start, x_end);
-      for (size_t j = 0; j < repeat_num; ++j) {
+      if (repeat_num > 0) {
+        auto x_sub_tensor = x->Slice(x_start, x_end);
+        x_sub_tensor.Resize({1, x_sub_tensor.numel()});
         int out_start = out_offset;
         if (x_lod.size() == 1) {
           out_start = out_lod[0][out_offset];
-          out_lod[0].push_back(x_seq_len);
         }
-        auto out_sub_tensor = out->Slice(out_start, out_start + x_seq_len);
-        framework::TensorCopy(x_sub_tensor, context.GetPlace(),
-                              &out_sub_tensor);
+        auto out_sub_tensor =
+            out->Slice(out_start, out_start + x_seq_len * repeat_num);
+        out_sub_tensor.Resize({repeat_num, x_sub_tensor.dims()[1]});
+        EigenMatrix<T>::From(out_sub_tensor).device(eigen_place) =
+            EigenMatrix<T>::From(x_sub_tensor)
+                .broadcast(Eigen::array<int, 2>({{repeat_num, 1}}));
+      }
+      for (int j = 0; j < repeat_num; ++j) {
+        if (x_lod.size() == 1) {
+          out_lod[0].push_back(out_lod[0].back() + x_seq_len);
+        }
         out_offset++;
       }
     }
@@ -122,6 +134,9 @@ class SequenceExpandGradKernel : public framework::OpKernel<T> {
 
     auto& dev_ctx = context.template device_context<DeviceContext>();
 
+    math::SetConstant<DeviceContext, T> set_zero;
+    set_zero(dev_ctx, g_x, static_cast<T>(0));
+
     int g_out_offset = 0;
     for (size_t i = 1; i < y_lod[ref_level].size(); ++i) {
       int repeat_num = y_lod[ref_level][i] - y_lod[ref_level][i - 1];
@@ -133,12 +148,11 @@ class SequenceExpandGradKernel : public framework::OpKernel<T> {
           x_end = x_lod[0][i];
         }
         int x_seq_len = x_end - x_start;
-        auto column = x_seq_len * x->dims()[1];
         auto g_x_sub = g_x->Slice(x_start, x_end);
-        g_x_sub = framework::ReshapeToMatrix(g_x_sub, column);
+        g_x_sub.Resize(flatten_to_1d(g_x_sub.dims()));
         int g_out_end = g_out_offset + repeat_num * x_seq_len;
         auto g_out_sub = g_out->Slice(g_out_offset, g_out_end);
-        g_out_sub = framework::ReshapeToMatrix(g_out_sub, column);
+        g_out_sub.Resize({repeat_num, g_x_sub.dims()[0]});
         math::ColwiseSum<DeviceContext, T> col_sum;
         col_sum(dev_ctx, g_out_sub, &g_x_sub);
         g_out_offset += repeat_num * x_seq_len;

python/paddle/fluid/layers/nn.py

@@ -1781,52 +1781,52 @@ def conv2d_transpose(input,
     return out
 
 
-def sequence_expand(x, y, name=None):
+def sequence_expand(x, y, ref_level=-1, name=None):
     """Sequence Expand Layer. This layer will expand the input variable **x**
-    according to LoD information of **y**. And the following examples will
-    explain how sequence_expand works:
+    according to specified level lod of **y**. Please note that lod level of
+    **x** is at most 1 and rank of **x** is at least 2. When rank of **x**
+    is greater than 2, then it would be viewed as a 2-D tensor.
+    Following examples will explain how sequence_expand works:
 
     .. code-block:: text
 
         * Case 1
             x is a LoDTensor:
-                x.lod = [[0,       2, 3],
-                         [0, 1,    3, 4]]
-                x.data = [a, b, c, d]
+                x.lod  = [[0,   2,        4]]
+                x.data = [[a], [b], [c], [d]]
                 x.dims = [4, 1]
 
             y is a LoDTensor:
                 y.lod = [[0,    2,    4],
                          [0, 3, 6, 7, 8]]
 
-            with condition len(y.lod[-1]) - 1 == x.dims[0]
+            ref_level: 0
 
-            then output is a 2-level LoDTensor:
-                out.lod = [[0,                2,    4],
-                           [0,       3,       6, 7, 8]]
-                out.data = [a, a, a, b, b, b, c, d]
+            then output is a 1-level LoDTensor:
+                out.lod =  [[0,   2,        4,        6,        8]]
+                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
                 out.dims = [8, 1]
 
         * Case 2
             x is a Tensor:
-                x.data = [a, b, c]
+                x.data = [[a], [b], [c]]
                 x.dims = [3, 1]
 
             y is a LoDTensor:
-                y.lod = [[0, 2, 3, 6]]
-
-            with condition len(y.lod[-1]) - 1 == x.dims[0]
+                y.lod = [[0, 2, 2, 5]]
 
-            then output is a 1-level LoDTensor:
-                out.lod = [[0,    2, 3,      6]]
-                out.data = [a, a, b, c, c, c]
-                out.dims = [6, 1]
+            ref_level: -1
 
+            then output is a Tensor:
+                out.data = [[a], [a], [c], [c], [c]]
+                out.dims = [5, 1]
     Args:
         x (Variable): The input variable which is a Tensor or LoDTensor.
         y (Variable): The input variable which is a LoDTensor.
+        ref_level (int): Lod level of `y` to be referred by `x`. If set to -1,
+                         refer the last level of lod.
         name(str|None): A name for this layer(optional). If set None, the layer
-                       will be named automatically.
+                        will be named automatically.
 
     Returns:
         Variable: The expanded variable which is a LoDTensor.
@@ -1837,14 +1837,17 @@ def sequence_expand(x, y, name=None):
             x = fluid.layers.data(name='x', shape=[10], dtype='float32')
             y = fluid.layers.data(name='y', shape=[10, 20],
                              dtype='float32', lod_level=1)
-            out = layers.sequence_expand(x=x, y=y)
+            out = layers.sequence_expand(x=x, y=y, ref_level=0)
     """
     helper = LayerHelper('sequence_expand', input=x, **locals())
     dtype = helper.input_dtype()
     tmp = helper.create_tmp_variable(dtype)
     helper.append_op(
-        type='sequence_expand', inputs={'X': x,
-                                        'Y': y}, outputs={'Out': tmp})
+        type='sequence_expand',
+        inputs={'X': x,
+                'Y': y},
+        outputs={'Out': tmp},
+        attrs={'ref_level': ref_level})
     return tmp
 
 

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

@@ -181,8 +181,8 @@ def test_sequence_expand(self):
         with program_guard(program):
             x = layers.data(name='x', shape=[10], dtype='float32')
             y = layers.data(
-                name='y', shape=[10, 20], dtype='float32', lod_level=1)
-            self.assertIsNotNone(layers.sequence_expand(x=x, y=y))
+                name='y', shape=[10, 20], dtype='float32', lod_level=2)
+            self.assertIsNotNone(layers.sequence_expand(x=x, y=y, ref_level=1))
         print(str(program))
 
     def test_lstm_unit(self):