m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #4061 from pkuyym/fix-4029

Browse Source 
Add expand operator
mobile_baidu
qingqing01 7 years ago committed by GitHub
parent 58b4c9af34 d7e7a1d7a5
commit 561d6340a4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 428 additions and 0 deletions
Show Stats Download Patch File Download Diff File

136
paddle/operators/expand_op.cc
Unescape View File

@ -0,0 +1,136 @@
/* 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/expand_op.h"
namespace paddle {
namespace operators {
using framework::Tensor;
class ExpandOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null.");
std::vector<int> expand_times =
ctx->Attrs().Get<std::vector<int>>("expand_times");
auto x_dims = ctx->GetInputDim("X");
PADDLE_ENFORCE_EQ(static_cast<size_t>(x_dims.size()), expand_times.size(),
"The number of Attr(expand_times)'s value must be equal "
"to the rank of Input(X).");
PADDLE_ENFORCE_LE(x_dims.size(), 6,
"The rank of Input(X) must not be greater than 6.");
std::vector<int64_t> out_shape(x_dims.size());
for (size_t i = 0; i < expand_times.size(); ++i) {
PADDLE_ENFORCE_GE(expand_times[i], 1,
"Each value of Attr(expand_times) should not be "
"less than 1.");
out_shape[i] = x_dims[i] * expand_times[i];
}
ctx->SetOutputDim("Out", framework::make_ddim(out_shape));
if (out_shape[0] == x_dims[0]) {
ctx->ShareLoD("X", "Out");
}
}
};
class ExpandOpMaker : public framework::OpProtoAndCheckerMaker {
public:
ExpandOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
"(Tensor, default Tensor<float>) A tensor with rank in [1, 6]."
"X is the input tensor to be expanded.");
AddOutput("Out",
"(Tensor, default Tensor<float>) A tensor with rank in [1, 6]."
"The rank of Output(Out) is same as Input(X) except that each "
"dimension size of Output(Out) is equal to corresponding "
"dimension size of Input(X) multiplying corresponding value of "
"Attr(expand_times).");
AddAttr<std::vector<int>>("expand_times",
"Expand times number for each dimension.");
AddComment(R"DOC(
Expand operator tiles the input by given times number. You should set times
number for each dimension by providing attribute 'expand_times'. The rank of X
should be in [1, 6]. Please notice that size of 'expand_times' must be same with
X's rank. Following is a using case:
Input(X) is a 3-D tensor with shape [2, 3, 1]:
[
[[1], [2], [3]],
[[4], [5], [6]]
]
Attr(expand_times): [1, 2, 2]
Output(Out) is a 3-D tensor with shape [2, 6, 2]:
[
[[1, 1], [2, 2], [3, 3], [1, 1], [2, 2], [3, 3]],
[[4, 4], [5, 5], [6, 6], [4, 4], [5, 5], [6, 6]]
]
)DOC");
}
};
class ExpandGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
"Input(Out@GRAD) should not be null.");
auto x_dims = ctx->GetInputDim("X");
std::vector<int> expand_times =
ctx->Attrs().Get<std::vector<int>>("expand_times");
auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
for (size_t i = 0; i < expand_times.size(); ++i) {
PADDLE_ENFORCE_EQ(x_dims[i] * expand_times[i], out_dims[i],
"Each dimension size of Input(Out@GRAD) should be "
"equal to multiplication of crroresponding dimension "
"size of Input(X) and Attr(expand_times) value.");
}
auto x_grad_name = framework::GradVarName("X");
if (ctx->HasOutput(x_grad_name)) {
ctx->SetOutputDim(x_grad_name, x_dims);
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(expand, ops::ExpandOp, ops::ExpandOpMaker, expand_grad,
ops::ExpandGradOp);
REGISTER_OP_CPU_KERNEL(expand,
ops::ExpandKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
expand_grad, ops::ExpandGradKernel<paddle::platform::CPUPlace, float>);

23
paddle/operators/expand_op.cu
Unescape View File

@ -0,0 +1,23 @@
/* 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/expand_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(expand,
ops::ExpandKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
expand_grad, ops::ExpandGradKernel<paddle::platform::GPUPlace, float>);

172
paddle/operators/expand_op.h
Unescape View File

@ -0,0 +1,172 @@
/* 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 <boost/preprocessor/arithmetic/div.hpp>
#include <boost/preprocessor/arithmetic/mod.hpp>
#include <boost/preprocessor/comparison/greater.hpp>
#include <boost/preprocessor/comparison/greater_equal.hpp>
#include <boost/preprocessor/control/if.hpp>
#include <boost/preprocessor/repetition/repeat.hpp>
#include <iostream>
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#define MAX_RANK_SUPPORTED 6
#define EXPAND_TEMPLATE(z, n, data) \
case n + 1: { \
Expand<n + 1>(context); \
break; \
}
#define REP_EXPAND_TEMPLATE(n) BOOST_PP_REPEAT(n, EXPAND_TEMPLATE, ~)
#define COND(n) \
BOOST_PP_GREATER_EQUAL(BOOST_PP_DIV(n, MAX_RANK_SUPPORTED), \
BOOST_PP_MOD(n, MAX_RANK_SUPPORTED))
#define EXPAND_GRAD_CASE(n) \
case n: { \
ExpandBackward<n>(context, reshape_dims_vec, reduce_dims_vec); \
break; \
}
#define EXPAND_GRAD_TEMPLATE(z, n, data) \
BOOST_PP_IF(COND(n), EXPAND_GRAD_CASE(n), )
#define REP_EXPAND_GRAD_TEMPLATE(n) BOOST_PP_REPEAT(n, EXPAND_GRAD_TEMPLATE, ~)
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename T, size_t D, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
template <typename Place, typename T>
class ExpandKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto rank = context.Input<Tensor>("X")->dims().size();
switch (rank) {
REP_EXPAND_TEMPLATE(MAX_RANK_SUPPORTED)
default:
PADDLE_ENFORCE(false,
"Only support tensor with rank being between 1 and 6.");
}
}
protected:
template <int Rank>
void Expand(const framework::ExecutionContext& context) const {
auto* in0 = context.Input<Tensor>("X");
auto& expand_times = context.Attr<std::vector<int>>("expand_times");
auto* out0 = context.Output<Tensor>("Out");
Eigen::DSizes<int, Rank> bcast_dims;
auto x_dims = in0->dims();
for (size_t i = 0; i < expand_times.size(); ++i) {
bcast_dims[i] = expand_times[i];
}
auto x = EigenTensor<T, Rank>::From(*in0);
out0->mutable_data<T>(context.GetPlace());
auto y = EigenTensor<T, Rank>::From(*out0);
auto place = context.GetEigenDevice<Place>();
y.device(place) = x.broadcast(bcast_dims);
}
};
template <typename Place, typename T>
class ExpandGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* in0 = context.Input<Tensor>("X");
auto& expand_times = context.Attr<std::vector<int>>("expand_times");
auto x_dims = in0->dims();
// 1. reshape_dims_vec is the broadcast parameter. For each dimension i,
// if expand_times[i] > 1 and x_dims[i] > 1, i will be splitted to two
// dimensions [expand_times[i], x_dims[i]].
// 2. reduce_dims_vec is the dimension parameter to compute gradients. For
// each dimension expanded, the gradients should be summed to original
// size.
std::vector<int> reshape_dims_vec;
std::vector<int> reduce_dims_vec;
for (size_t i = 0; i < expand_times.size(); ++i) {
if (expand_times[i] == 1) {
reshape_dims_vec.push_back(x_dims[i]);
} else {
if (x_dims[i] == 1) {
reduce_dims_vec.push_back(reshape_dims_vec.size());
reshape_dims_vec.push_back(expand_times[i]);
} else {
reduce_dims_vec.push_back(reshape_dims_vec.size());
reshape_dims_vec.push_back(expand_times[i]);
reshape_dims_vec.push_back(x_dims[i]);
}
}
}
int dims = reshape_dims_vec.size() * MAX_RANK_SUPPORTED +
reduce_dims_vec.size() - MAX_RANK_SUPPORTED - 1;
// no need reduce, just copy
if (reduce_dims_vec.size() == 0) {
auto* in0 = context.Input<Tensor>(framework::GradVarName("Out"));
auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
out0->mutable_data<T>(context.GetPlace());
out0->CopyFrom(*in0, context.GetPlace(), context.device_context());
} else {
switch (dims) {
REP_EXPAND_GRAD_TEMPLATE(72)
default:
PADDLE_ENFORCE(
false, "Only support tensor with rank being between 1 and 6.");
}
}
}
protected:
template <int Dims>
void ExpandBackward(const framework::ExecutionContext& context,
const std::vector<int>& reshape_dims_vec,
const std::vector<int>& reduce_dims_vec) const {
size_t reshape_size = Dims / MAX_RANK_SUPPORTED + 1;
size_t reduce_size = Dims % MAX_RANK_SUPPORTED + 1;
PADDLE_ENFORCE_EQ(reshape_size, reshape_dims_vec.size(),
"Inconsistent size between template Dims and "
"reshape dimensions.");
PADDLE_ENFORCE_EQ(reduce_size, reduce_dims_vec.size(),
"Inconsistent size between template Dims and "
"reduce dimensions.");
auto* in0 = context.Input<Tensor>(framework::GradVarName("Out"));
auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
auto x = EigenVector<T>::Flatten(*(context.Input<Tensor>("X")));
out0->mutable_data<T>(context.GetPlace());
auto x_grad = EigenVector<T>::Flatten(*out0);
Eigen::DSizes<int, Dims / MAX_RANK_SUPPORTED + 1> reshape_dims;
for (size_t i = 0; i < reshape_size; ++i) {
reshape_dims[i] = reshape_dims_vec[i];
}
Eigen::DSizes<int, Dims % MAX_RANK_SUPPORTED + 1> reduce_dims;
for (size_t i = 0; i < reduce_size; ++i) {
reduce_dims[i] = reduce_dims_vec[i];
}
auto out_grad = EigenVector<T>::Flatten(*in0);
x_grad.device(context.GetEigenDevice<Place>()) =
out_grad.reshape(reshape_dims).sum(reduce_dims).reshape(x.dimensions());
}
};
} // namespace operators
} // namespace paddle

97
python/paddle/v2/framework/tests/test_expand_op.py
Unescape View File

@ -0,0 +1,97 @@
import unittest
import numpy as np
from op_test import OpTest
class TestExpandOpRank1(OpTest):
def setUp(self):
self.op_type = "expand"
self.inputs = {'X': np.random.random(12).astype("float32")}
self.attrs = {'expand_times': [2]}
output = np.tile(self.inputs['X'], 2)
self.outputs = {'Out': output}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
class TestExpandOpRank2_Corner(OpTest):
def setUp(self):
self.op_type = "expand"
self.inputs = {'X': np.random.random((12, 14)).astype("float32")}
self.attrs = {'expand_times': [1, 1]}
output = np.tile(self.inputs['X'], (1, 1))
self.outputs = {'Out': output}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
class TestExpandOpRank2(OpTest):
def setUp(self):
self.op_type = "expand"
self.inputs = {'X': np.random.random((12, 14)).astype("float32")}
self.attrs = {'expand_times': [2, 3]}
output = np.tile(self.inputs['X'], (2, 3))
self.outputs = {'Out': output}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
class TestExpandOpRank3_Corner(OpTest):
def setUp(self):
self.op_type = "expand"
self.inputs = {'X': np.random.random((2, 4, 5)).astype("float32")}
self.attrs = {'expand_times': [1, 1, 1]}
output = np.tile(self.inputs['X'], (1, 1, 1))
self.outputs = {'Out': output}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
class TestExpandOpRank3(OpTest):
def setUp(self):
self.op_type = "expand"
self.inputs = {'X': np.random.random((2, 4, 5)).astype("float32")}
self.attrs = {'expand_times': [2, 1, 4]}
output = np.tile(self.inputs['X'], (2, 1, 4))
self.outputs = {'Out': output}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
class TestExpandOpRank4(OpTest):
def setUp(self):
self.op_type = "expand"
self.inputs = {'X': np.random.random((2, 4, 5, 7)).astype("float32")}
self.attrs = {'expand_times': [3, 2, 1, 2]}
output = np.tile(self.inputs['X'], (3, 2, 1, 2))
self.outputs = {'Out': output}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
if __name__ == "__main__":
unittest.main()
Write Preview
Loading…
Cancel
Save