提交 26857659 编写于 作者: Q Qiao Longfei 提交者: GitHub

add cross-entropy-op (#2965)

* add cross-entropy-op

* add infershape and compute

* implement Infershape and compute of onehotcrossentropy op
上级 c90b94e8
...@@ -48,6 +48,7 @@ op_library(mul_op SRCS mul_op.cc mul_op.cu) ...@@ -48,6 +48,7 @@ op_library(mul_op SRCS mul_op.cc mul_op.cu)
op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc) op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
op_library(sigmoid_op SRCS sigmoid_op.cu sigmoid_op.cc) op_library(sigmoid_op SRCS sigmoid_op.cu sigmoid_op.cc)
op_library(softmax_op SRCS softmax_op.cc softmax_op.cu) op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu)
op_library(fc_op SRCS fc_op.cc DEPS mul_op rowwise_add_op sigmoid_op op_library(fc_op SRCS fc_op.cc DEPS mul_op rowwise_add_op sigmoid_op
softmax_op net) softmax_op net)
......
/* 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/cross_entropy_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor.h"
namespace paddle {
namespace operators {
class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 2,
"Input size of OnehotCrossEntropyOp must be two");
PADDLE_ENFORCE(outputs.size() == 1,
"Output size of OnehotCrossEntropyOp must be one");
PADDLE_ENFORCE(inputs[0] != nullptr && inputs[1] != nullptr,
"Inputs of OnehotCrossEntropyOp must all be set");
PADDLE_ENFORCE(outputs[0] != nullptr,
"Outputs of OnehotCrossEntropyOp must all be set");
PADDLE_ENFORCE(inputs[0]->dims().size() == 2, "X's dimension must be 2.");
PADDLE_ENFORCE(outputs[0]->dims().size() == 1,
"label's dimension must be 1.");
outputs[0]->set_dims(framework::make_ddim({inputs[0]->dims()[0]}));
}
};
class OnehotCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
public:
OnehotCrossEntropyOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of OnehotCrossEntropyOp");
AddInput("label", "The second input of OnehotCrossEntropyOp");
AddOutput("Y", "The output of OnehotCrossEntropyOp");
AddComment(R"DOC(
OnehotCrossEntropy Operator.
Y[i] = -log(X[i][j])
)DOC");
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP(onehot_cross_entropy,
paddle::operators::OnehotCrossEntropyOp,
paddle::operators::OnehotCrossEntropyOpMaker);
REGISTER_OP_CPU_KERNEL(
onehot_cross_entropy,
paddle::operators::OnehotCrossEntropyOpKernel<::paddle::platform::CPUPlace,
float>);
#include "paddle/operators/cross_entropy_op.h"
#include "paddle/framework/op_registry.h"
REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
paddle::operators::OnehotCrossEntropyOpKernel<
::paddle::platform::GPUPlace, float>);
\ No newline at end of file
/* 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 "glog/logging.h"
#include "paddle/framework/operator.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class OnehotCrossEntropyOpKernel : public framework::OpKernel {
public:
constexpr T LOG_THRESHOLD() const { return static_cast<T>(1e-20); }
void Compute(const framework::KernelContext& context) const override {
auto X = context.Input(0)->Get<framework::Tensor>();
const T* X_data = X.data<T>();
const int* label_data =
context.Input(1)->Get<framework::Tensor>().data<int>();
auto* Y = context.Output(0)->GetMutable<framework::Tensor>();
Y->mutable_data<T>(context.GetPlace());
T* Y_data = Y->data<T>();
int batch_size = X.dims()[0];
int class_num = X.dims()[1];
// Y[i] = -log(X[i][j])
for (int i = 0; i < batch_size; ++i) {
Y_data[i] = -std::log(
std::max(X_data[i * class_num + label_data[i]], LOG_THRESHOLD()));
}
}
};
} // namespace operators
} // namespace paddle
cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python
add_op fc_op sgd_op) add_op fc_op sgd_op cross_entropy_op)
...@@ -27,6 +27,7 @@ namespace py = pybind11; ...@@ -27,6 +27,7 @@ namespace py = pybind11;
namespace pd = paddle::framework; namespace pd = paddle::framework;
USE_OP(add_two); USE_OP(add_two);
USE_OP(onehot_cross_entropy);
USE_OP_WITHOUT_KERNEL(fc); USE_OP_WITHOUT_KERNEL(fc);
USE_OP(sgd); USE_OP(sgd);
......
add_python_test(test_framework test_protobuf.py test_scope.py add_python_test(test_framework test_protobuf.py test_scope.py
test_default_scope_funcs.py test_op_creation_methods.py test_default_scope_funcs.py test_op_creation_methods.py
test_tensor.py test_fc_op.py test_add_two_op.py test_sgd_op.py) test_tensor.py test_fc_op.py test_add_two_op.py test_sgd_op.py test_cross_entropy_op.py)
import unittest
import numpy
from op_test_util import OpTestMeta
class TestSGD(unittest.TestCase):
__metaclass__ = OpTestMeta
def setUp(self):
self.type = "onehot_cross_entropy"
batch_size = 100
class_num = 10
self.X = numpy.random.random((batch_size, class_num)).astype("float32")
self.label = 5 * numpy.ones(batch_size).astype("int32")
Y = []
for i in range(0, batch_size):
Y.append(-numpy.log(self.X[i][self.label[i]]))
self.Y = numpy.array(Y).astype("float32")
if __name__ == "__main__":
unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册