root
/
MNN
mirror of https://github.com/alibaba/MNN.git
1
0
Fork 0
Code Issues Actions 1 Packages Projects Releases Wiki Activity
MNN/tools/converter/source/optimizer/merge/Conv1dQuantToConv2dQuant.cpp

313 lines
10 KiB
C++
Raw Blame History

//
// Conv1dQuantToConv2dQuant.cpp
// MNNConverter
//
// Created by MNN on 2020/08/10.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "../TemplateMerge.hpp"
#include "MNN_generated.h"
namespace MNN {
namespace Express {
static auto gRegister = []() {
auto match = [](EXPRP expr) {
if (nullptr == expr->get()) {
return false;
}
if (expr->get()->type() != OpType::OpType_ConvertTensor) {
return false;
}
auto input1 = expr->inputs()[0];
auto expr1 = input1->expr().first;
if (expr1->get() == nullptr) {
return false;
}
if (expr1->get()->type() != OpType::OpType_Int8ToFloat) {
return false;
}
auto input2 = expr1->inputs()[0];
auto expr2 = input2->expr().first;
if (expr2->get() == nullptr) {
return false;
}
if (expr2->get()->type() != OpType::OpType_FloatToInt8) {
return false;
}
auto input3 = expr2->inputs()[0];
auto expr3 = input3->expr().first;
if (expr3->get() == nullptr) {
return false;
}
if (expr3->get()->type() != OpType::OpType_ConvertTensor) {
return false;
}
auto input4 = expr3->inputs()[0];
auto expr4 = input4->expr().first;
if (expr4->get() == nullptr) {
return false;
}
auto input5 = expr4->inputs()[0];
auto expr5 = input5->expr().first;
if (expr5->get() == nullptr) {
return false;
}
auto input6 = expr5->inputs()[0];
auto expr6 = input6->expr().first;
if (expr6->get() == nullptr) {
return false;
}
int squeezeIndex = 0;
if (expr4->get()->type() == OpType::OpType_Squeeze) {
squeezeIndex = 0;
} else if (expr5->get()->type() == OpType::OpType_Squeeze) {
squeezeIndex = 1;
} else if (expr6->get()->type() == OpType::OpType_Squeeze) {
squeezeIndex = 2;
} else {
return false; // should find squeeze in 3 steps after OpType_ConvertTensor
}
EXPRP squeezeExpr = nullptr;
if (squeezeIndex == 0) {
squeezeExpr = expr4;
}
if (squeezeIndex == 1) {
squeezeExpr = expr5;
if (expr4->get()->type() == OpType_BinaryOp) {
auto binaryOp = expr4->get();
auto binaryParams = binaryOp->main_as_BinaryOp();
if (binaryParams->opType() != BinaryOpOperation_ADD) {
return false;
}
} else if ((expr4->get()->type() != OpType_ReLU) && (expr4->get()->type() != OpType_ReLU6)) {
return false;
} else {
// pass
}
}
if (squeezeIndex == 2) {
squeezeExpr = expr6;
if ((expr4->get()->type() != OpType_ReLU) && (expr4->get()->type() != OpType_ReLU6)) {
return false;
}
if (expr5->get()->type() != OpType_BinaryOp) {
return false;
}
auto binaryOp = expr5->get();
auto binaryParams = binaryOp->main_as_BinaryOp();
if (binaryParams->opType() != BinaryOpOperation_ADD) {
return false;
}
}
auto input7 = squeezeExpr->inputs()[0];
auto expr7 = input7->expr().first;
if (expr7->get() == nullptr) {
return false;
}
if (expr7->get()->type() != OpType::OpType_Convolution) {
return false;
}
auto input8 = expr7->inputs()[0];
auto expr8 = input8->expr().first;
if (expr8->get() == nullptr) {
return false;
}
if (expr8->get()->type() != OpType_ConvertTensor) {
return false;
}
auto input9 = expr8->inputs()[0];
auto expr9 = input9->expr().first;
if (expr9->get() == nullptr) {
return false;
}
if (expr9->get()->type() != OpType_Int8ToFloat) {
return false;
}
auto input10 = expr9->inputs()[0];
auto expr10 = input10->expr().first;
if (expr10->get() == nullptr) {
return false;
}
if (expr10->get()->type() != OpType_FloatToInt8) {
return false;
}
auto input11 = expr10->inputs()[0];
auto expr11 = input11->expr().first;
if (expr11->get() == nullptr) {
return false;
}
if (expr11->get()->type() != OpType_ConvertTensor) {
return false;
}
auto input12 = expr11->inputs()[0];
auto expr12 = input12->expr().first;
if (expr12->get() == nullptr) {
return false;
}
if (expr12->get()->type() != OpType_ExpandDims) {
return false;
}
return true;
};
auto transform = [](EXPRP expr) {
// OpType_Int8ToFloat
auto input1 = expr->inputs()[0];
auto expr1 = input1->expr().first;
// OpType_FloatToInt8
auto input2 = expr1->inputs()[0];
auto expr2 = input2->expr().first;
// OpType_ConvertTensor
auto input3 = expr2->inputs()[0];
auto expr3 = input3->expr().first;
auto input4 = expr3->inputs()[0];
auto expr4 = input4->expr().first;
auto input5 = expr4->inputs()[0];
auto expr5 = input5->expr().first;
auto input6 = expr5->inputs()[0];
auto expr6 = input6->expr().first;
int squeezeIndex = 0;
if (expr4->get()->type() == OpType::OpType_Squeeze) {
squeezeIndex = 0;
} else if (expr5->get()->type() == OpType::OpType_Squeeze) {
squeezeIndex = 1;
} else if (expr6->get()->type() == OpType::OpType_Squeeze) {
squeezeIndex = 2;
} else {
// should find squeeze in 3 steps after OpType_ConvertTensor
}
EXPRP squeezeExpr = nullptr;
if (squeezeIndex == 0) {
squeezeExpr = expr4;
}
if (squeezeIndex == 1) {
squeezeExpr = expr5;
// then, expr4 should be relu or relu6 or bias_add
}
if (squeezeIndex == 2) {
squeezeExpr = expr6;
// then, expr4 should be relu or relu6
// expr5 should be bias_add
}
// OpType_Convolution
auto input7 = squeezeExpr->inputs()[0];
auto expr7 = input7->expr().first;
// OpType_ConvertTensor
auto input8 = expr7->inputs()[0];
auto expr8 = input8->expr().first;
// OpType_Int8ToFloat
auto input9 = expr8->inputs()[0];
auto expr9 = input9->expr().first;
// OpType_FloatToInt8
auto input10 = expr9->inputs()[0];
auto expr10 = input10->expr().first;
// OpType_ConvertTensor
auto input11 = expr10->inputs()[0];
auto expr11 = input11->expr().first;
// OpType_ExpandDims
auto input12 = expr11->inputs()[0];
auto expr12 = input12->expr().first;
// now, begin reorder
auto convInput = expr7->inputs()[0];
std::unique_ptr<OpT> op7(expr7->get()->UnPack());
auto newExpr7 = Expr::create(op7.get(), {convInput});
newExpr7->setName(expr7->name());
auto output = Variable::create(newExpr7);
output->setName(expr7->outputName(0));
if (squeezeIndex == 0) {
// pass
}
if (squeezeIndex == 1) {
std::vector<VARP> inputs{output};
if (expr4->get()->type() == OpType_BinaryOp) {
inputs.push_back(expr4->inputs().at(1));
}
std::unique_ptr<OpT> op4(expr4->get()->UnPack());
auto newExpr4 = Expr::create(op4.get(), inputs);
newExpr4->setName(expr4->name());
output = Variable::create(newExpr4);
output->setName(expr4->outputName(0));
}
if (squeezeIndex == 2) {
std::unique_ptr<OpT> op5(expr5->get()->UnPack());
auto newExpr5 = Expr::create(op5.get(), {output, expr5->inputs().at(1)});
newExpr5->setName(expr5->name());
output = Variable::create(newExpr5);
output->setName(expr5->outputName(0));
std::unique_ptr<OpT> op4(expr4->get()->UnPack());
auto newExpr4 = Expr::create(op4.get(), {output});
newExpr4->setName(expr4->name());
output = Variable::create(newExpr4);
output->setName(expr4->outputName(0));
}
std::unique_ptr<OpT> op3(expr3->get()->UnPack());
auto newExpr3 = Expr::create(op3.get(), {output});
newExpr3->setName(expr3->name());
output = Variable::create(newExpr3);
output->setName(expr3->outputName(0));
std::unique_ptr<OpT> op2(expr2->get()->UnPack());
auto newExpr2 = Expr::create(op2.get(), {output});
newExpr2->setName(expr2->name());
output = Variable::create(newExpr2);
output->setName(expr2->outputName(0));
std::unique_ptr<OpT> op1(expr1->get()->UnPack());
auto newExpr1 = Expr::create(op1.get(), {output});
newExpr1->setName(expr1->name());
output = Variable::create(newExpr1);
output->setName(expr1->outputName(0));
std::unique_ptr<OpT> op(expr->get()->UnPack());
auto newExpr = Expr::create(op.get(), {output});
newExpr->setName(expr->name());
output = Variable::create(newExpr);
output->setName(expr->outputName(0));
std::unique_ptr<OpT> squeezeOp(squeezeExpr->get()->UnPack());
auto newSqueezeExpr = Expr::create(squeezeOp.get(), {output});
newSqueezeExpr->setName(squeezeExpr->name());
output = Variable::create(newSqueezeExpr);
output->setName(squeezeExpr->outputName(0));
Expr::replace(expr, output->expr().first);
return true;
};
TemplateMerge::getInstance("Merge").insertTemplate("Conv1dQuantToConv2dQuant", match, transform,
PASS_PRIORITY_HIGH);
return true;
}();
}
} // namespace MNN
Reference in New Issue View Git Blame Copy Permalink