wehub-resource-sync
37 lines
1.4 KiB
C++
37 lines
1.4 KiB
C++
//
|
|
// TransformOnnxPad.cpp
|
|
// MNNConverter
|
|
//
|
|
// Created by MNN on 2019/09/05.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "../PostTreatUtils.hpp"
|
|
using namespace MNN;
|
|
class TransformOnnxPad : public PostConverter {
|
|
public:
|
|
virtual bool onExecute(std::unique_ptr<MNN::NetT>& net) const override {
|
|
for (auto iter = net->oplists.begin(); iter != net->oplists.end();) {
|
|
auto op = iter->get();
|
|
if (OpType_Padding == op->type && op->main.type == OpParameter_Blob && op->inputIndexes.size() == 1) {
|
|
std::unique_ptr<OpT> paddingConst(new OpT);
|
|
paddingConst->type = OpType_Const;
|
|
paddingConst->main.type = OpParameter_Blob;
|
|
paddingConst->main.value = new BlobT(*op->main.AsBlob());
|
|
paddingConst->name = op->name + "padding";
|
|
paddingConst->outputIndexes = {(int)net->tensorName.size()};
|
|
net->tensorName.emplace_back(paddingConst->name);
|
|
op->inputIndexes = {op->inputIndexes[0], paddingConst->outputIndexes[0]};
|
|
op->main.Reset();
|
|
iter = net->oplists.insert(iter, std::move(paddingConst));
|
|
iter++;
|
|
iter++;
|
|
continue;
|
|
}
|
|
iter++;
|
|
}
|
|
return true;
|
|
}
|
|
};
|
|
static PostConverterRegister<TransformOnnxPad> __l("TransformOnnxPad");
|