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

Merge pull request #4071 from QiJune/activation_ops

Browse Source 
Implement activation related operators
update-doc-pybind
QI JUN 8 years ago committed by GitHub
parent d827359c71 87ba6cbf20
commit 8c3b8af31e
9 changed files with 989 additions and 174 deletions
Show Stats Download Patch File Download Diff File

7
paddle/operators/CMakeLists.txt
Unescape View File

@ -55,6 +55,13 @@ function(op_library TARGET)
set(pybind_flag 1)
endif()
# activation_op contains several operators
if ("${TARGET}" STREQUAL "activation_op")
set(pybind_flag 1)
# It's enough to just adding one operator to pybind
file(APPEND ${pybind_file} "USE_OP(sigmoid);\n")
endif()
# pybind USE_NO_KERNEL_OP
file(READ ${TARGET}.cc TARGET_CONTENT)
string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")

306
paddle/operators/activation_op.cc
View File

File diff suppressed because it is too large Load Diff

100
paddle/operators/activation_op.cu
Unescape View File

@ -0,0 +1,100 @@
/* 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/activation_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(sigmoid,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::SigmoidFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
sigmoid_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::SigmoidGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
exp,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::ExpFunctor>);
REGISTER_OP_GPU_KERNEL(exp_grad,
ops::ActivationGradKernel<paddle::platform::GPUPlace,
float, ops::ExpGradFunctor>);
REGISTER_OP_GPU_KERNEL(relu,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::ReluFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
relu_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::ReluGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
tanh,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::TanhFunctor>);
REGISTER_OP_GPU_KERNEL(
tanh_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::TanhGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
sqrt,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::SqrtFunctor>);
REGISTER_OP_GPU_KERNEL(
sqrt_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::SqrtGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
abs,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::AbsFunctor>);
REGISTER_OP_GPU_KERNEL(abs_grad,
ops::ActivationGradKernel<paddle::platform::GPUPlace,
float, ops::AbsGradFunctor>);
REGISTER_OP_GPU_KERNEL(reciprocal,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::ReciprocalFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
reciprocal_grad,
ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::ReciprocalGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
log,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::LogFunctor>);
REGISTER_OP_GPU_KERNEL(
log_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::LogGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(square,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::SquareFunctor>);
REGISTER_OP_GPU_KERNEL(
square_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::SquareGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(brelu,
ops::BReluKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(brelu_grad,
ops::BReluGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(soft_relu,
ops::SoftReluKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
soft_relu_grad, ops::SoftReluGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(pow, ops::PowKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(pow_grad,
ops::PowGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(stanh,
ops::STanhKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(stanh_grad,
ops::STanhGradKernel<paddle::platform::GPUPlace, float>);

353
paddle/operators/activation_op.h
View File

File diff suppressed because it is too large Load Diff

67
paddle/operators/sigmoid_op.cc
Unescape View File

@ -1,67 +0,0 @@
/* 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/sigmoid_op.h"
namespace paddle {
namespace operators {
class SigmoidOp : 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 SigmoidOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
"Output(Y) of SigmoidOp should not be null.");
ctx.Output<framework::LoDTensor>("Y")->Resize(
ctx.Input<Tensor>("X")->dims());
}
};
class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SigmoidOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "sigmoid input");
AddOutput("Y", "sigmoid output");
AddComment("Sigmoid function");
}
};
class SigmoidOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
->Resize(ctx.Input<Tensor>("Y")->dims());
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad,
ops::SigmoidOpGrad);
REGISTER_OP_CPU_KERNEL(sigmoid,
ops::SigmoidKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
sigmoid_grad, ops::SigmoidGradKernel<paddle::platform::CPUPlace, float>);

23
paddle/operators/sigmoid_op.cu
Unescape View File

@ -1,23 +0,0 @@
/* 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/sigmoid_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(sigmoid,
ops::SigmoidKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
sigmoid_grad, ops::SigmoidGradKernel<paddle::platform::GPUPlace, float>);

62
paddle/operators/sigmoid_op.h
Unescape View File

@ -1,62 +0,0 @@
/* 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"
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 Place, typename T>
class SigmoidKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto input = context.Input<Tensor>("X");
auto output = context.Output<Tensor>("Y");
output->mutable_data<T>(context.GetPlace());
// The clipping is used in Paddle's raw implenmention
auto X = EigenVector<T>::Flatten(*input);
auto Y = EigenVector<T>::Flatten(*output);
auto place = context.GetEigenDevice<Place>();
Y.device(place) = 1. / (1. + (-X).exp());
}
};
template <typename Place, typename T>
class SigmoidGradKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto Y_t = context.Input<Tensor>("Y");
auto dY_t = context.Input<Tensor>(framework::GradVarName("Y"));
auto dX_t = context.Output<Tensor>(framework::GradVarName("X"));
dX_t->mutable_data<T>(context.GetPlace());
auto dX = EigenVector<T>::Flatten(*dX_t);
auto Y = EigenVector<T>::Flatten(*Y_t);
auto dY = EigenVector<T>::Flatten(*dY_t);
dX.device(context.GetEigenDevice<Place>()) = dY * Y * (1. - Y);
}
};
} // namespace operators
} // namespace paddle

223
python/paddle/v2/framework/tests/test_activation_op.py
Unescape View File

@ -0,0 +1,223 @@
import unittest
import numpy as np
from op_test import OpTest
class TestExp(OpTest):
def setUp(self):
self.op_type = "exp"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': np.exp(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestSigmoid(OpTest):
def setUp(self):
self.op_type = "sigmoid"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.008)
class TestTanh(OpTest):
def setUp(self):
self.op_type = "tanh"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': np.tanh(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestSqrt(OpTest):
def setUp(self):
self.op_type = "sqrt"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': np.sqrt(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestAbs(OpTest):
def setUp(self):
self.op_type = "abs"
x = np.random.uniform(-1, 1, [4, 4]).astype("float32")
# Because we set delta = 0.005 in caculating numeric gradient,
# if x is too small, such as 0.002, x_neg will be -0.003
# x_pos will be 0.007, so the numeric gradient is unaccurate.
# we should avoid this
x[np.abs(x) < 0.005] = 0.02
self.inputs = {'X': x}
self.outputs = {'Y': np.abs(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestRelu(OpTest):
def setUp(self):
self.op_type = "relu"
x = np.random.uniform(-1, 1, [11, 17]).astype("float32")
# The same reason with TestAbs
x[np.abs(x) < 0.005] = 0.02
self.inputs = {'X': x}
self.outputs = {'Y': np.maximum(self.inputs['X'], 0)}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestBRelu(OpTest):
def setUp(self):
self.op_type = "brelu"
x = np.random.uniform(-1, 1, [4, 4]).astype("float32")
t_min = 1
t_max = 4
# The same with TestAbs
x[np.abs(x - t_min) < 0.005] = t_min + 0.02
x[np.abs(x - t_max) < 0.005] = t_max + 0.02
self.inputs = {'X': x}
self.attrs = {'t_min': t_min, 't_max': t_max}
t = np.copy(x)
t[t < t_min] = t_min
t[t > t_max] = t_max
self.outputs = {'Y': t}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.02)
class TestSoftRelu(OpTest):
def setUp(self):
self.op_type = "soft_relu"
x = np.random.uniform(-3, 3, [4, 4]).astype("float32")
threshold = 2
# The same reason with TestAbs
x[np.abs(x - threshold) < 0.005] = threshold + 0.02
x[np.abs(x + threshold) < 0.005] = -threshold + 0.02
self.inputs = {'X': x}
self.attrs = {'threshold': threshold}
t = np.copy(x)
t[t < -threshold] = -threshold
t[t > threshold] = threshold
self.outputs = {'Y': np.log((np.exp(t) + 1))}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.02)
class TestReciprocal(OpTest):
def setUp(self):
self.op_type = "reciprocal"
self.inputs = {'X': np.random.uniform(1, 2, [11, 17]).astype("float32")}
self.outputs = {'Y': np.reciprocal(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.01)
class TestLog(OpTest):
def setUp(self):
self.op_type = "log"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': np.log(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestSquare(OpTest):
def setUp(self):
self.op_type = "square"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': np.square(self.inputs['X'])}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
class TestPow(OpTest):
def setUp(self):
self.op_type = "pow"
self.inputs = {'X': np.random.uniform(1, 2, [11, 17]).astype("float32")}
self.attrs = {'factor': 3}
self.outputs = {'Y': np.power(self.inputs['X'], 3)}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.02)
class TestSTanh(OpTest):
def setUp(self):
self.op_type = "stanh"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
scale_a = 2.0 / 3.0
scale_b = 1.7159
self.attrs = {'scale_a': scale_a, 'scale_b': scale_b}
self.outputs = {'Y': scale_b * np.tanh(self.inputs['X'] * scale_a)}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Y', max_relative_error=0.007)
if __name__ == "__main__":
unittest.main()

22
python/paddle/v2/framework/tests/test_sigmoid_op.py
Unescape View File

@ -1,22 +0,0 @@
import unittest
import numpy as np
from op_test import OpTest
class TestSigmoidOp(OpTest):
def setUp(self):
self.op_type = "sigmoid"
self.inputs = {
'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
}
self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(["X"], "Y", max_relative_error=0.007)
if __name__ == '__main__':
unittest.main()
Write Preview
Loading…
Cancel
Save