wehub-resource-sync
216 lines
7.8 KiB
C++
216 lines
7.8 KiB
C++
//
|
|
// OpGrad.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/05/05.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include <mutex>
|
|
#include "OpGrad.hpp"
|
|
using namespace std;
|
|
using namespace MNN::Express;
|
|
//#define MNN_TRAIN_DEBUG
|
|
namespace MNN {
|
|
extern void registerGradOps();
|
|
static std::map<int, OpGrad*>& getConverter() {
|
|
static std::map<int, OpGrad*> gConverterMap;
|
|
return gConverterMap;
|
|
}
|
|
|
|
OpGrad* OpGrad::get(int type) {
|
|
auto& converterMap = getConverter();
|
|
auto iter = converterMap.find(type);
|
|
if (iter != converterMap.end()) {
|
|
return iter->second;
|
|
}
|
|
return nullptr;
|
|
}
|
|
|
|
void OpGrad::insert(int type, OpGrad* converter) {
|
|
auto& converterMap = getConverter();
|
|
converterMap.insert(std::make_pair(type, converter));
|
|
}
|
|
|
|
std::vector<Express::VARP> OpGrad::gradLinear(Express::VARP loss, const std::vector<Express::VARP>& parameters, const std::vector<Express::VARP>& outputDiff, const std::vector<std::string> blockExpr) {
|
|
std::map<EXPRP, std::vector<VARP>> backwardMap;
|
|
auto outputSize = loss->expr().first->outputSize();
|
|
if (outputSize != outputDiff.size()) {
|
|
MNN_ERROR("The expr output %d, but diff size is %d\n", outputSize, (int)outputDiff.size());
|
|
return {};
|
|
}
|
|
backwardMap[loss->expr().first] = outputDiff;
|
|
std::set<VARP> parameterSet;
|
|
for (auto p : parameters) {
|
|
parameterSet.insert(p);
|
|
}
|
|
auto res = gradCommon({loss}, parameterSet, backwardMap, blockExpr);
|
|
std::vector<VARP> linearRes(parameters.size(), nullptr);
|
|
for (int i=0; i<parameters.size(); ++i) {
|
|
auto iter = res.find(parameters[i]);
|
|
if (iter != res.end()) {
|
|
linearRes[i] = iter->second;
|
|
}
|
|
}
|
|
return linearRes;
|
|
}
|
|
|
|
std::map<Express::VARP, Express::VARP> OpGrad::grad(VARP loss, const std::set<Express::VARP>& parameters, const std::vector<std::string> blockName) {
|
|
std::map<EXPRP, std::vector<VARP>> backwardMap;
|
|
{
|
|
auto shape = loss->getInfo();
|
|
MNN_ASSERT(shape->size == 1);
|
|
auto init = _Const(1.0f, shape->dim, shape->order);
|
|
backwardMap[loss->expr().first] = std::vector<VARP>{init};
|
|
}
|
|
return gradCommon({loss}, parameters, backwardMap, blockName);
|
|
}
|
|
Express::VARP OpGrad::divideAvoidZero(MNN::Express::VARP y, MNN::Express::VARP x) {
|
|
auto p = MNN::Express::_Abs(x);
|
|
auto sx = MNN::Express::_Sign(x);
|
|
p = MNN::Express::_Maximum(p, MNN::Express::_Scalar<float>(0.000001f));
|
|
return MNN::Express::_Divide(y, p) * sx;
|
|
}
|
|
static std::once_flag gInit;
|
|
void OpGrad::init() {
|
|
std::call_once(gInit, []() {
|
|
registerGradOps();
|
|
});
|
|
}
|
|
|
|
std::pair<std::vector<Express::VARP>, std::vector<Express::VARP>> OpGrad::gradCommon(std::vector<Express::VARP> outputs, std::vector<Express::VARP> outputDiff, std::vector<Express::VARP> parameters) {
|
|
if (outputs.size() != outputDiff.size()) {
|
|
MNN_ERROR("outputDiff size %d not equal output size %d\n", (int)outputs.size(), (int)outputDiff.size());
|
|
return {};
|
|
}
|
|
std::map<EXPRP, std::vector<VARP>> backwardMap;
|
|
for (int i=0; i<outputs.size(); ++i) {
|
|
auto expr = outputs[i]->expr();
|
|
if (backwardMap.find(expr.first) == backwardMap.end()) {
|
|
std::vector<Express::VARP> res(expr.first->outputSize(), nullptr);
|
|
backwardMap.insert(std::make_pair(expr.first, res));
|
|
}
|
|
auto iter = backwardMap.find(expr.first);
|
|
if (nullptr == iter->second[expr.second]) {
|
|
iter->second[expr.second] = outputDiff[i];
|
|
} else {
|
|
iter->second[expr.second] = iter->second[expr.second] + outputDiff[i];
|
|
}
|
|
}
|
|
std::set<Express::VARP> parameterSets;
|
|
for (auto p : parameters) {
|
|
parameterSets.insert(p);
|
|
}
|
|
auto varmap = gradCommon(outputs, parameterSets, backwardMap);
|
|
std::vector<Express::VARP> res;
|
|
std::vector<Express::VARP> resDiff;
|
|
for (int i=0; i<parameters.size(); ++i) {
|
|
auto iter = varmap.find(parameters[i]);
|
|
if (iter != varmap.end()) {
|
|
res.push_back(iter->first);
|
|
resDiff.push_back(iter->second);
|
|
}
|
|
}
|
|
return std::make_pair(res, resDiff);
|
|
}
|
|
|
|
std::map<Express::VARP, Express::VARP> OpGrad::gradCommon(std::vector<Express::VARP> outputs, const std::set<Express::VARP>& parameters, std::map<EXPRP, std::vector<VARP>>& backwardMap, const std::vector<std::string> blockName) {
|
|
init();
|
|
auto executeOrder = Variable::getExecuteOrder(outputs);
|
|
for (auto iter = executeOrder.rbegin(); iter != executeOrder.rend(); iter++) {
|
|
auto expr = *iter;
|
|
auto& inputs = expr->inputs();
|
|
if (backwardMap.find(expr) == backwardMap.end()) {
|
|
continue;
|
|
}
|
|
if (nullptr == expr->get()) {
|
|
continue;
|
|
}
|
|
auto grad = OpGrad::get(expr->get()->type());
|
|
#ifdef MNN_TRAIN_DEBUG
|
|
MNN_PRINT("Grad for %s, %s\n", expr->name().c_str(), MNN::EnumNameOpType(expr->get()->type()));
|
|
#endif
|
|
if (nullptr == grad) {
|
|
#ifdef MNN_TRAIN_DEBUG
|
|
MNN_PRINT("Can't grad for %s, %s\n", expr->name().c_str(), MNN::EnumNameOpType(expr->get()->type()));
|
|
#endif
|
|
continue;
|
|
}
|
|
auto inputGrad = grad->onGrad(expr, backwardMap[expr]);
|
|
if (!expr->name().empty()) {
|
|
for (int v=0; v<inputGrad.size(); ++v) {
|
|
if (inputGrad[v].get() != nullptr) {
|
|
inputGrad[v]->setName("grad::" + expr->name() + std::to_string(v));
|
|
}
|
|
}
|
|
}
|
|
auto empty = true;
|
|
for (auto grad : inputGrad) {
|
|
if (nullptr != grad) {
|
|
empty = false;
|
|
break;
|
|
}
|
|
}
|
|
if (empty) {
|
|
#ifdef MNN_TRAIN_DEBUG
|
|
MNN_PRINT("Can't grad for %s, %d\n", expr->name().c_str(), expr->get()->type());
|
|
#endif
|
|
continue;
|
|
}
|
|
if (!blockName.empty()) {
|
|
if (std::find(blockName.begin(), blockName.end(), expr->name()) != blockName.end()) {
|
|
for (int ii = 0; ii <inputGrad.size(); ii++) {
|
|
inputGrad[ii] = nullptr;
|
|
}
|
|
continue;
|
|
}
|
|
}
|
|
#ifdef MNN_TRAIN_DEBUG
|
|
for (int i = 0; i < inputGrad.size(); ++i) {
|
|
if (nullptr == inputGrad[i]) {
|
|
continue;
|
|
}
|
|
auto info = inputGrad[i]->getInfo();
|
|
if (nullptr == info) {
|
|
MNN_ERROR("Grad error for %s, %d\n", expr->name().c_str(), expr->get()->type());
|
|
break;
|
|
}
|
|
}
|
|
#endif
|
|
MNN_ASSERT(inputGrad.size() <= inputs.size());
|
|
for (int i = 0; i < inputGrad.size(); ++i) {
|
|
auto inputExpr = inputs[i]->expr().first;
|
|
auto index = inputs[i]->expr().second;
|
|
auto backward = inputGrad[i];
|
|
if (nullptr == backward) {
|
|
continue;
|
|
}
|
|
if (backwardMap.find(inputExpr) == backwardMap.end()) {
|
|
backwardMap.insert(std::make_pair(inputExpr, std::vector<VARP>(inputExpr->outputSize())));
|
|
}
|
|
auto& inputVarMap = backwardMap[inputExpr];
|
|
if (nullptr == inputVarMap[index]) {
|
|
inputVarMap[index] = backward;
|
|
} else {
|
|
inputVarMap[index] = _Add(inputVarMap[index], backward);
|
|
}
|
|
}
|
|
}
|
|
std::map<Express::VARP, Express::VARP> grads;
|
|
std::map<Expr*, VARP> parametersExpr;
|
|
for (auto p : parameters) {
|
|
parametersExpr.insert(std::make_pair(p->expr().first.get(), p));
|
|
}
|
|
for (auto iter : backwardMap) {
|
|
auto expr = iter.first.get();
|
|
if (parametersExpr.find(expr) != parametersExpr.end()) {
|
|
auto parameter = parametersExpr[expr];
|
|
grads[parameter] = iter.second[parameter->expr().second];
|
|
}
|
|
}
|
|
// MNN_PRINT("Grad: %d <- %d\n", grads.size(), parameters.size());
|
|
return grads;
|
|
}
|
|
|
|
} // namespace MNN
|