wehub-resource-sync
71 lines
2.8 KiB
C++
71 lines
2.8 KiB
C++
//
|
|
// Loss.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/11/29.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "Loss.hpp"
|
|
|
|
using namespace MNN::Express;
|
|
|
|
namespace MNN {
|
|
namespace Train {
|
|
|
|
Express::VARP _CrossEntropy(Express::VARP predicts, Express::VARP oneHotTargets) {
|
|
MNN_ASSERT(predicts->getInfo()->dim.size() == 2);
|
|
MNN_ASSERT(predicts->getInfo()->dim == oneHotTargets->getInfo()->dim);
|
|
auto loss = _Negative(_ReduceMean(_ReduceSum(_Log(predicts) * oneHotTargets, {1}), {}));
|
|
return loss;
|
|
}
|
|
|
|
Express::VARP _KLDivergence(Express::VARP predicts, Express::VARP oneHotTargets) {
|
|
MNN_ASSERT(predicts->getInfo()->dim.size() == 2);
|
|
MNN_ASSERT(predicts->getInfo()->dim == oneHotTargets->getInfo()->dim);
|
|
auto loss = _ReduceMean(_ReduceSum(_Multiply(predicts, _Log(predicts) - _Log(oneHotTargets)), {1}), {});
|
|
return loss;
|
|
}
|
|
|
|
Express::VARP _MSE(Express::VARP predicts, Express::VARP oneHotTargets) {
|
|
MNN_ASSERT(predicts->getInfo()->dim.size() == 2);
|
|
MNN_ASSERT(predicts->getInfo()->dim == oneHotTargets->getInfo()->dim);
|
|
auto loss = _ReduceMean(_ReduceSum(_Square(predicts - oneHotTargets), {1}), {});
|
|
return loss;
|
|
}
|
|
|
|
Express::VARP _MAE(Express::VARP predicts, Express::VARP oneHotTargets) {
|
|
MNN_ASSERT(predicts->getInfo()->dim.size() == 2);
|
|
MNN_ASSERT(predicts->getInfo()->dim == oneHotTargets->getInfo()->dim);
|
|
auto loss = _ReduceMean(_ReduceSum(_Abs(predicts - oneHotTargets), {1}), {});
|
|
return loss;
|
|
}
|
|
|
|
Express::VARP _Hinge(Express::VARP predicts, Express::VARP oneHotTargets) {
|
|
MNN_ASSERT(predicts->getInfo()->dim.size() == 2);
|
|
MNN_ASSERT(predicts->getInfo()->dim == oneHotTargets->getInfo()->dim);
|
|
auto loss = _ReduceMean(_ReduceSum(_Maximum(_Const(0.), _Const(1.) - predicts * oneHotTargets), {1}), {});
|
|
return loss;
|
|
}
|
|
|
|
Express::VARP _DistillLoss(Express::VARP studentLogits, Express::VARP teacherLogits, Express::VARP oneHotTargets, const float temperature, const float alpha) {
|
|
auto info = teacherLogits->getInfo();
|
|
if (info->order == NC4HW4) {
|
|
teacherLogits = _Convert(teacherLogits, NCHW);
|
|
studentLogits = _Convert(studentLogits, NCHW);
|
|
}
|
|
MNN_ASSERT(studentLogits->getInfo()->dim.size() == 2);
|
|
MNN_ASSERT(studentLogits->getInfo()->dim == teacherLogits->getInfo()->dim);
|
|
MNN_ASSERT(studentLogits->getInfo()->dim == oneHotTargets->getInfo()->dim);
|
|
MNN_ASSERT(alpha >= 0 && alpha <= 1);
|
|
auto softTargets = _Softmax(teacherLogits * _Scalar(1 / temperature));
|
|
auto studentPredict = _Softmax(studentLogits * _Scalar(1 / temperature));
|
|
auto loss1 = _Scalar(temperature * temperature) * _KLDivergence(studentPredict, softTargets);
|
|
auto loss2 = _CrossEntropy(_Softmax(studentLogits), oneHotTargets);
|
|
auto loss = _Scalar(alpha) * loss1 + _Scalar(1 - alpha) * loss2;
|
|
return loss;
|
|
}
|
|
|
|
} // namespace Train
|
|
} // namespace MNN
|