From f14a7966b0f982a703f24940b1e9ae5325ee83c9 Mon Sep 17 00:00:00 2001 From: Liu Yiqun Date: Thu, 21 Sep 2017 03:18:42 +0000 Subject: [PATCH] Initialize the sequence softmax operator. --- paddle/operators/sequence_avg_pool_op.cc | 2 +- paddle/operators/sequence_softmax_op.cc | 82 +++++++++++++++++++ paddle/operators/sequence_softmax_op.cu | 25 ++++++ paddle/operators/sequence_softmax_op.h | 62 ++++++++++++++ .../tests/test_sequence_softmax_op.py | 35 ++++++++ 5 files changed, 205 insertions(+), 1 deletion(-) create mode 100644 paddle/operators/sequence_softmax_op.cc create mode 100644 paddle/operators/sequence_softmax_op.cu create mode 100644 paddle/operators/sequence_softmax_op.h create mode 100644 python/paddle/v2/framework/tests/test_sequence_softmax_op.py diff --git a/paddle/operators/sequence_avg_pool_op.cc b/paddle/operators/sequence_avg_pool_op.cc index 9815b8f3a8d..ff354c62b6f 100644 --- a/paddle/operators/sequence_avg_pool_op.cc +++ b/paddle/operators/sequence_avg_pool_op.cc @@ -36,7 +36,7 @@ class SequenceAvgPoolOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_GE( dims[0], /*batch size = */ static_cast(lod[0].size() - 1), - "The first dimension of Input(X) must be large than batch size."); + "The first dimension of Input(X) must be larger than batch size."); dims[0] = lod[0].size() - 1; ctx.Output("Out")->Resize({dims}); } diff --git a/paddle/operators/sequence_softmax_op.cc b/paddle/operators/sequence_softmax_op.cc new file mode 100644 index 00000000000..0a99717440e --- /dev/null +++ b/paddle/operators/sequence_softmax_op.cc @@ -0,0 +1,82 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/sequence_softmax_op.h" + +namespace paddle { +namespace operators { + +class SequenceSoftmaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_NOT_NULL( + ctx.InputVar("X"), "Input(X) of SequenceSoftmaxOp should not be null."); + PADDLE_ENFORCE_NOT_NULL( + ctx.OutputVar("Out"), + "Output(Out) of SequenceSoftmaxOp should not be null."); + + auto *x = ctx.Input("X"); + auto dims = x->dims(); + auto lod = x->lod(); + PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); + PADDLE_ENFORCE_GE( + dims[0], + /* batch_size */ static_cast(lod[0].size() - 1), + "The first dimension of Input(X) should be larger than batch size."); + PADDLE_ENFORCE_EQ(x->numel(), static_cast(lod[0].size() - 1), + "The width of each timestep in Input(X) of " + "SequenceSoftmaxOp should be 1."); + + dims[0] = lod[0].size() - 1; + ctx.Output("Out")->Resize({dims}); + } +}; + +class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SequenceSoftmaxOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(LoDTensor)"); + AddOutput("Out", "(LoDTensor)"); + AddComment(R"DOC( +Softmax of Sequence. +)DOC"); + } +}; + +class SequenceSoftmaxGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sequence_softmax, ops::SequenceSoftmaxOp, + ops::SequenceSoftmaxOpMaker, sequence_softmax_grad, + ops::SequenceSoftmaxGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_softmax, + ops::SequenceSoftmaxKernel); +REGISTER_OP_CPU_KERNEL( + sequence_softmax_grad, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.cu b/paddle/operators/sequence_softmax_op.cu new file mode 100644 index 00000000000..f2a1e3d5e31 --- /dev/null +++ b/paddle/operators/sequence_softmax_op.cu @@ -0,0 +1,25 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/sequence_softmax_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + sequence_softmax, + ops::SequenceSoftmaxKernel) +REGISTER_OP_GPU_KERNEL( + sequence_softmax_grad, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.h b/paddle/operators/sequence_softmax_op.h new file mode 100644 index 00000000000..54d82652711 --- /dev/null +++ b/paddle/operators/sequence_softmax_op.h @@ -0,0 +1,62 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/softmax_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; +template +using EigenVector = framework::EigenVector; +template +using EigenMatrix = framework::EigenMatrix; + +template +class SequenceSoftmaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + + auto lod = x->lod(); + const size_t level = lod.size(); + + out->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + Tensor x_i = x->Slice(start_pos, end_pos); + Tensor out_i = out->Slice(start_pos, end_pos); + + math::SoftmaxFunctor()(&x_i, &out_i, ctx); + } + } +}; + +template +class SequenceSoftmaxGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override {} +}; + +} // namespace operators +} // namespace paddle diff --git a/python/paddle/v2/framework/tests/test_sequence_softmax_op.py b/python/paddle/v2/framework/tests/test_sequence_softmax_op.py new file mode 100644 index 00000000000..d0667c1308d --- /dev/null +++ b/python/paddle/v2/framework/tests/test_sequence_softmax_op.py @@ -0,0 +1,35 @@ +import unittest +import numpy as np +from op_test import OpTest + + +def stable_softmax(x): + """Compute the softmax of vector x in a numerically stable way.""" + shiftx = x - np.max(x) + exps = np.exp(shiftx) + return exps / np.sum(exps) + + +class TestSequenceSoftmaxOp(OpTest): + def setUp(self): + self.op_type = "sequence_softmax" + x = np.random.uniform(0.1, 1, (11, 1)).astype("float32") + lod = [[0, 4, 5, 8, 11]] + + out = np.zeros((11, 1)).astype("float32") + for i in range(4): + sub_x = x[lod[0][i]:lod[0][i + 1], :] + sub_x = sub_x.reshape(1, lod[0][i + 1] - lod[0][i]) + sub_out = stable_softmax(sub_x) + out[lod[0][i]:lod[0][i + 1], :] = sub_out.reshape( + lod[0][i + 1] - lod[0][i], 1) + + self.inputs = {"X": (x, lod)} + self.outputs = {"Out": out} + + def test_check_output(self): + self.check_output() + + +if __name__ == "__main__": + unittest.main() -- GitLab