wehub-resource-sync
53 lines
1.6 KiB
C++
53 lines
1.6 KiB
C++
//
|
|
// MVNTransform.cpp
|
|
// MNNConverter
|
|
//
|
|
// Created by MNN on 2019/12/12.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include <vector>
|
|
#include "CaffeExtraManager.hpp"
|
|
#include "MNN_generated.h"
|
|
namespace MNN {
|
|
namespace Express {
|
|
|
|
class MVNTransform : public CaffeExtraManager::Transform {
|
|
public:
|
|
virtual EXPRP onExecute(EXPRP expr) const override {
|
|
auto op = expr->get();
|
|
auto inputs = expr->inputs();
|
|
auto acrossChannels = op->main_as_Extra()->attr()->GetAs<Attribute>(0)->b();
|
|
auto eps = op->main_as_Extra()->attr()->GetAs<Attribute>(1)->f();
|
|
auto normalizeVariance = op->main_as_Extra()->attr()->GetAs<Attribute>(2)->b();
|
|
|
|
std::vector<int> reduceDims;
|
|
if (acrossChannels) {
|
|
reduceDims = {1, 2, 3};
|
|
} else {
|
|
reduceDims = {2, 3};
|
|
}
|
|
|
|
auto mean = _ReduceMean(inputs[0], reduceDims, true);
|
|
auto subMean = _Subtract(inputs[0], mean); // of input shape
|
|
|
|
if (!normalizeVariance) {
|
|
return subMean->expr().first;
|
|
} else {
|
|
auto s2 = _Square(subMean); // element wise of input shape
|
|
auto variance = _ReduceMean(s2, reduceDims, true);
|
|
auto stdv = _Add(_Sqrt(variance), _Const(eps));
|
|
auto normedData = _Divide(inputs[0], stdv);
|
|
return normedData->expr().first;
|
|
}
|
|
}
|
|
};
|
|
|
|
static auto gRegister = []() {
|
|
CaffeExtraManager::get()->insert("MVN", std::shared_ptr<CaffeExtraManager::Transform>(new MVNTransform));
|
|
return true;
|
|
}();
|
|
|
|
} // namespace Express
|
|
} // namespace MNN
|