diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index 35d251f71a0cb631d5900498ea3188b5ddeae334..17e576a9d5c8f50fbe84b066a93460f03ae6bb08 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -371,6 +371,8 @@ template struct RowwiseAdd; template struct ColwiseSum; template struct ColwiseSum; +template struct ColwiseSum; +template struct ColwiseSum; template struct RowwiseSum; template struct RowwiseSum; diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu index 3abbcdb71d03eaf6f8eba3d97150d27ac5a5405e..c6ca2693a053360ce5dc44765acf1520a11cce2c 100644 --- a/paddle/fluid/operators/math/math_function.cu +++ b/paddle/fluid/operators/math/math_function.cu @@ -422,6 +422,8 @@ struct RowwiseAdd { template struct RowwiseAdd; template struct RowwiseAdd; template struct ColwiseSum; +template struct ColwiseSum; +template struct ColwiseSum; // template struct ColwiseSum; // The ColwiseSum failed in debug mode, // and only failed for this case. So reimplemented it. diff --git a/paddle/fluid/operators/sequence_expand_op.cc b/paddle/fluid/operators/sequence_expand_op.cc index a5d84d629b2e50763dac9bc571ac490414a8a406..786fe63e7580ce16b946d5049a490eed2c3c6ced 100644 --- a/paddle/fluid/operators/sequence_expand_op.cc +++ b/paddle/fluid/operators/sequence_expand_op.cc @@ -17,7 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { -using framework::Tensor; +using framework::LoDTensor; class SequenceExpandOp : public framework::OperatorWithKernel { public: @@ -25,15 +25,71 @@ class SequenceExpandOp : public framework::OperatorWithKernel { protected: void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X")); - PADDLE_ENFORCE(ctx->HasOutput("Out")); - PADDLE_ENFORCE(ctx->HasInput("Y")); - framework::DDim out_dim; - auto y_dim = ctx->GetInputDim("Y"); - out_dim = ctx->GetInputDim("X"); - out_dim[0] = y_dim[0]; - ctx->ShareLoD("Y", "Out"); - ctx->SetOutputDim("Out", out_dim); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceExpandOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of SequenceExpandOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "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("ref_level"); + + PADDLE_ENFORCE_GE(x_dims.size(), 2, + "Dimension number of Input(X) should be at least 2."); + + if (ctx->IsRuntime()) { + framework::Variable* x_var = + boost::get(ctx->GetInputVarPtrs("X")[0]); + framework::Variable* y_var = + boost::get(ctx->GetInputVarPtrs("Y")[0]); + + auto& x_lod = x_var->Get().lod(); + auto& y_lod = y_var->Get().lod(); + + PADDLE_ENFORCE_LE(x_lod.size(), 1, + "Level number of Input(X)'s lod should not be " + "greater than 1."); + 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(y_lod.size())), + "Invlid `ref_level`, which should be either equal to -1 " + "or in [0, %d)", + y_lod.size()); + + if (ref_level == -1) ref_level = y_lod.size() - 1; + + if (x_lod.size() > 0) { + PADDLE_ENFORCE(x_lod[0].size() == y_lod[ref_level].size(), + "Level number of Input(X)'s lod could be 0. Otherwise " + "size of Input(X)'s first level lod should be equal to " + "size of Input(Y)'s referred level lod."); + } + + int64_t out_first_dim = 0; + if (y_lod[ref_level].size() <= 1) { + out_first_dim = x_dims[0]; + } else { + for (size_t i = 1; i < y_lod[ref_level].size(); ++i) { + int x_seq_len = 1; + if (x_lod.size() == 1) { + x_seq_len = x_lod[0][i] - x_lod[0][i - 1]; + } + out_first_dim += + (y_lod[ref_level][i] - y_lod[ref_level][i - 1]) * x_seq_len; + } + } + out_dims[0] = out_first_dim; + ctx->SetOutputDim("Out", out_dims); + } else { + out_dims[0] = -1; + ctx->SetOutputDim("Out", out_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } } }; @@ -42,83 +98,81 @@ class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker { SequenceExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", - "(Tensor or LoDTensor) The input(X) of this operator can be a " - "LoDTensor or a base Tensor."); + "(LoDTensor, default LoDTensor) A 2-D LoDTensor whose lod " + "level is at most 1."); AddInput("Y", - "(LoDTensor)The reference input(Y) of sequence_expand op." - "It must be a LoDTensor with k-level(k>0)." - "The input(X) will be expanded according to LOD of input(Y)." - "The element numbers of last level in input(Y) " - "must be equal to dims[0] of input(X)."); + "(LoDTensor, default LoDTensor) Referred LoDTensor whose " + "lod (specified level) is referred by Input(X)."); AddOutput("Out", - "(LodTensor)The output of sequence_expand op." - "The lod of output will be as same as input(Y)'s lod."); + "(LodTensor, default LoDTensor) Output LoDTensor which is " + "generated from Input(X) by referring lod of Input(Y)."); + AddAttr("ref_level", "Specify lod level of Input(Y).").SetDefault(-1); 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, 1, 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 get 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"); } }; @@ -129,12 +183,14 @@ class SequenceExpandOpGrad : public framework::OperatorWithKernel { protected: void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X")); - PADDLE_ENFORCE(ctx->HasInput("Out")); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) should not be null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), - "The input(Out@GRAD) should not be null"); + "Input(Out@GRAD) should not be null."); + auto x_dims = ctx->GetInputDim("X"); auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { ctx->SetOutputDim(x_grad_name, x_dims); } @@ -149,7 +205,13 @@ REGISTER_OP(sequence_expand, ops::SequenceExpandOp, ops::SequenceExpandOpMaker, sequence_expand_grad, ops::SequenceExpandOpGrad); REGISTER_OP_CPU_KERNEL( sequence_expand, - ops::SequenceExpandKernel); + ops::SequenceExpandKernel, + ops::SequenceExpandKernel, + ops::SequenceExpandKernel, + ops::SequenceExpandKernel); REGISTER_OP_CPU_KERNEL( sequence_expand_grad, - ops::SequenceExpandGradKernel); + ops::SequenceExpandGradKernel, + ops::SequenceExpandGradKernel, + ops::SequenceExpandGradKernel, + ops::SequenceExpandGradKernel); diff --git a/paddle/fluid/operators/sequence_expand_op.cu b/paddle/fluid/operators/sequence_expand_op.cu index 26622d23afa1c703e237628bcb11db8f1da73210..bb51bb2902eea797de3449fcb6c8b52b4f0e7fbf 100644 --- a/paddle/fluid/operators/sequence_expand_op.cu +++ b/paddle/fluid/operators/sequence_expand_op.cu @@ -18,7 +18,14 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( sequence_expand, - ops::SequenceExpandKernel); + ops::SequenceExpandKernel, + ops::SequenceExpandKernel, + ops::SequenceExpandKernel, + ops::SequenceExpandKernel); REGISTER_OP_CUDA_KERNEL( sequence_expand_grad, - ops::SequenceExpandGradKernel); + ops::SequenceExpandGradKernel, + ops::SequenceExpandGradKernel, + ops::SequenceExpandGradKernel, + ops::SequenceExpandGradKernel); diff --git a/paddle/fluid/operators/sequence_expand_op.h b/paddle/fluid/operators/sequence_expand_op.h index 76dde976db2d19e307ae7406be8280f9b4987187..db7d8bd6821fabd9714a160970558291ec47197f 100644 --- a/paddle/fluid/operators/sequence_expand_op.h +++ b/paddle/fluid/operators/sequence_expand_op.h @@ -16,45 +16,75 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/memory/memcpy.h" -#include "unsupported/Eigen/CXX11/Tensor" +#include "paddle/fluid/operators/math/math_function.h" namespace paddle { namespace operators { using LoDTensor = framework::LoDTensor; +template +using EigenMatrix = framework::EigenMatrix; template class SequenceExpandKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); - auto* out = context.Output("Out"); - const T* x_data = x->data(); - auto x_dims = x->dims(); auto* y = context.Input("Y"); - PADDLE_ENFORCE(!y->lod().empty(), "y should have lod"); - PADDLE_ENFORCE_EQ(static_cast(x_dims[0]), - y->lod().back().size() - 1, - "The size of last lod level in Input(Y)" - "must be equal to dims[0] of Input(X)."); - out->set_lod(y->lod()); - auto* place = - context.template device_context().eigen_device(); - size_t element_len = framework::product(x_dims) / x_dims[0]; - T* out_data = out->mutable_data(context.GetPlace()); - auto out_starts = out->lod().back(); - - for (size_t i = 0; i < out_starts.size() - 1; i++) { - int scale = out_starts[i + 1] - out_starts[i]; - Eigen::TensorMap< - Eigen::Tensor> - x_t(x_data, 1, element_len); - Eigen::TensorMap> - out_t(out_data, scale, element_len); - Eigen::array cast({{scale, 1}}); - out_t.device(*place) = x_t.broadcast(cast); - x_data += element_len; - out_data += element_len * scale; + auto* out = context.Output("Out"); + + int ref_level = context.Attr("ref_level"); + auto& x_lod = x->lod(); + auto& y_lod = y->lod(); + + if (ref_level == -1) ref_level = y_lod.size() - 1; + + out->mutable_data(context.GetPlace()); + + if (y_lod[ref_level].size() <= 1) { + framework::TensorCopy(*x, context.GetPlace(), out); + return; + } + + auto& out_lod = *out->mutable_lod(); + if (x_lod.size() == 1) { + out_lod.resize(1); + out_lod[0] = {0}; + } + + int out_offset = 0; + auto& eigen_place = + *context.template device_context().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; + int x_end = i; + if (x_lod.size() == 1) { + x_start = x_lod[0][i - 1]; + x_end = x_lod[0][i]; + } + int x_seq_len = x_end - x_start; + 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]; + } + 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::From(out_sub_tensor).device(eigen_place) = + EigenMatrix::From(x_sub_tensor) + .broadcast(Eigen::array({{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++; + } } } }; @@ -75,27 +105,51 @@ template class SequenceExpandGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* d_out = context.Input(framework::GradVarName("Out")); + auto* g_out = context.Input(framework::GradVarName("Out")); auto* x = context.Input("X"); - auto* out = context.Input("Out"); - auto* d_x = context.Output(framework::GradVarName("X")); - auto out_last_level = out->lod().back(); - d_x->set_lod(x->lod()); - const T* d_out_data = d_out->data(); - T* d_x_data = d_x->mutable_data(context.GetPlace()); - size_t element_len = d_out->numel() / d_out->dims()[0]; - for (size_t i = 0; i < out_last_level.size() - 1; ++i) { - size_t repeat = out_last_level[i + 1] - out_last_level[i]; - Eigen::TensorMap< - Eigen::Tensor> - d_out_t(d_out_data, static_cast(repeat), element_len); - Eigen::TensorMap> - d_x_t(d_x_data, static_cast(element_len)); - auto place = - context.template device_context().eigen_device(); - d_x_t.device(*place) = d_out_t.sum(Eigen::array({{0}})); - d_out_data += (repeat * element_len); - d_x_data += element_len; + auto* y = context.Input("Y"); + auto* g_x = context.Output(framework::GradVarName("X")); + int ref_level = context.Attr("ref_level"); + + g_x->mutable_data(context.GetPlace()); + g_x->set_lod(x->lod()); + + auto& x_lod = x->lod(); + auto& y_lod = y->lod(); + + if (ref_level == -1) ref_level = y_lod.size() - 1; + + // just copy the gradient + if (y_lod[ref_level].size() <= 1) { + framework::TensorCopy(*g_out, context.GetPlace(), g_x); + return; + } + + auto& dev_ctx = context.template device_context(); + + math::SetConstant set_zero; + set_zero(dev_ctx, g_x, static_cast(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]; + if (repeat_num > 0) { + int x_start = i - 1; + int x_end = i; + if (x_lod.size() == 1) { + x_start = x_lod[0][i - 1]; + x_end = x_lod[0][i]; + } + int x_seq_len = x_end - x_start; + auto g_x_sub = g_x->Slice(x_start, x_end); + 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.Resize({repeat_num, g_x_sub.dims()[0]}); + math::ColwiseSum col_sum; + col_sum(dev_ctx, g_out_sub, &g_x_sub); + g_out_offset += repeat_num * x_seq_len; + } } } }; diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 9656dcf94f14ad9250bb7e79c1330c9bdd44d9d6..75d3d895081e29e25fd5cf29d19e4b8459035ffb 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -1809,52 +1809,52 @@ 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]] + y.lod = [[0, 2, 2, 5]] - with condition len(y.lod[-1]) - 1 == x.dims[0] - - 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. @@ -1865,14 +1865,17 @@ 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 diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index fa38bd3762423497b82c3b421b3a1db4cd87525b..3a1a0859ecfd4ac5337e2112f8b22e32d8474f22 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -118,12 +118,12 @@ def decoder_decode(context, is_sparse): is_sparse=is_sparse) # use rnn unit to update rnn - current_state = pd.fc(input=[pre_ids_emb, pre_state_expanded], + current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb], size=decoder_size, act='tanh') - + current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score) # use score to do beam search - current_score = pd.fc(input=current_state, + current_score = pd.fc(input=current_state_with_lod, size=target_dict_dim, act='softmax') topk_scores, topk_indices = pd.topk(current_score, k=50) diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 744a762ae7664f1f28713c505f9112ba712fd41d..b5fd59cf3a1bea50b799c3ace8f3b9cea088b9d5 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -181,8 +181,8 @@ class TestBook(unittest.TestCase): 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): diff --git a/python/paddle/fluid/tests/unittests/test_sequence_expand.py b/python/paddle/fluid/tests/unittests/test_sequence_expand.py index 957fa5d2c4a795cfd01047c1b7845674e4c1d549..7feb509c4d6f5768552fc2515081f7e68f420967 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_expand.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_expand.py @@ -27,12 +27,36 @@ class TestSequenceExpand(OpTest): def compute(self): x = self.inputs['X'] x_data, x_lod = x if type(x) == tuple else (x, None) - n = 1 + x_data.shape[0] if not x_lod else len(x_lod[0]) y_data, y_lod = self.inputs['Y'] - repeats = [((y_lod[-1][i + 1] - y_lod[-1][i])) - for i in range(len(y_lod[-1]) - 1)] - out = x_data.repeat(repeats, axis=0) - self.outputs = {'Out': out} + + if hasattr(self, 'attrs'): + ref_level = self.attrs['ref_level'] + else: + ref_level = len(y_lod) - 1 + + out = np.zeros(shape=((0, ) + x_data.shape[1:]), dtype=x_data.dtype) + + if x_lod is None: + x_idx = [i for i in xrange(x_data.shape[0] + 1)] + else: + x_idx = x_lod[0] + out_lod = [[0]] + + for i in xrange(1, len(y_lod[ref_level])): + repeat_num = y_lod[ref_level][i] - y_lod[ref_level][i - 1] + x_len = x_idx[i] - x_idx[i - 1] + if repeat_num > 0: + x_sub = x_data[x_idx[i - 1]:x_idx[i], :] + x_sub = np.repeat(x_sub, repeat_num, axis=0) + out = np.vstack((out, x_sub)) + if x_lod is not None: + for j in xrange(repeat_num): + out_lod[0].append(out_lod[0][-1] + x_len) + + if x_lod is None: + self.outputs = {'Out': out} + else: + self.outputs = {'Out': (out, out_lod)} def setUp(self): self.op_type = 'sequence_expand' @@ -52,7 +76,8 @@ class TestSequenceExpandCase1(TestSequenceExpand): x_lod = [[0, 2, 5]] y_data = np.random.uniform(0.1, 1, [13, 1]).astype('float32') y_lod = [[0, 2, 5], [0, 2, 4, 7, 10, 13]] - self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} + self.inputs = {'X': x_data, 'Y': (y_data, y_lod)} + self.attrs = {'ref_level': 0} class TestSequenceExpandCase2(TestSequenceExpand): @@ -60,8 +85,9 @@ class TestSequenceExpandCase2(TestSequenceExpand): x_data = np.random.uniform(0.1, 1, [1, 2, 2]).astype('float32') x_lod = [[0, 1]] y_data = np.random.uniform(0.1, 1, [2, 2, 2]).astype('float32') - y_lod = [[0, 2]] + y_lod = [[0, 2], [0, 2]] self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} + self.attrs = {'ref_level': 0} class TestSequenceExpandCase3(TestSequenceExpand): @@ -75,14 +101,9 @@ class TestSequenceExpandCase3(TestSequenceExpand): class TestSequenceExpandCase4(TestSequenceExpand): def set_data(self): - x_data = np.array( - [0.1, 0.3, 0.2, 0.15, 0.25, 0.2, 0.15, 0.25, 0.1, 0.3]).reshape( - [2, 5]).astype('float32') - x_lod = [[ - 0, - 1, - 2, - ]] + data = [0.1, 0.3, 0.2, 0.15, 0.25, 0.2, 0.15, 0.25, 0.1, 0.3] + x_data = np.array(data).reshape([5, 2]).astype('float32') + x_lod = [[0, 2, 5]] y_data = np.random.uniform(0.1, 1, [2, 1]).astype('float32') y_lod = [[0, 1, 2], [0, 1, 2]] self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}