Caffe實(shí)踐C++源碼解讀:走入Solver

在理解如何使caffe運(yùn)行之后，我們要理解它是如何運(yùn)行的，即了解Solver類的Solve()函數(shù)做了什么，對(duì)于Solver類中如何初始化網(wǎng)絡(luò)以及其他參數(shù)，有興趣的可以深入研究。

源碼中Solver()函數(shù)是有參數(shù)形式的

??//?The?main?entry?of?the?solver?function.?In?default,?iter?will?be?zero.?Pass
??//?in?a?non-zero?iter?number?to?resume?training?for?a?pre-trained?net.
??virtual?void?Solve(const?char*?resume_file?=?NULL);
??inline?void?Solve(const?string?resume_file)?{?Solve(resume_file.c_str());?}

各位一看就明白了吧。再看Solve函數(shù)的定義

templatevoid?Solver::Solve(const?char*?resume_file)?{
??CHECK(Caffe::root_solver());
??LOG(INFO)?<<?"Solving?"?<<?net_->name();
??LOG(INFO)?<<?"Learning?Rate?Policy:?"?<<?param_.lr_policy();

??//?Initialize?to?false?every?time?we?start?solving.
??requested_early_exit_?=?false;

??if?(resume_file)?{
????LOG(INFO)?<<?"Restoring?previous?solver?status?from?"?<<?resume_file;
????Restore(resume_file);
??}

傳入?yún)?shù)resume_file是用于繼續(xù)中斷的訓(xùn)練的，既然caffe.cpp中的train()函數(shù)中已經(jīng)進(jìn)行了此操作，此處就不需要再傳入resume_file參數(shù)了。
然后Solver就直接進(jìn)入訓(xùn)練模式了，即Step函數(shù)，傳入?yún)?shù)為循環(huán)的次數(shù)，此參數(shù)在solver.txt文件中定義的max_iter和resume_file加載的iter_參數(shù)的差。

??//?For?a?network?that?is?trained?by?the?solver,?no?bottom?or?top?vecs
??//?should?be?given,?and?we?will?just?provide?dummy?vecs.
??int?start_iter?=?iter_;
??Step(param_.max_iter()?-?iter_);

在進(jìn)入Step函數(shù)之前，我們繼續(xù)往下看，訓(xùn)練完成后caffe會(huì)保存當(dāng)前模型

??//?If?we?haven't?already,?save?a?snapshot?after?optimization,?unless
??//?overridden?by?setting?snapshot_after_train?:=?false
??if?(param_.snapshot_after_train()
??????&&?(!param_.snapshot()?||?iter_?%?param_.snapshot()?!=?0))?{
????Snapshot();
??}

如果solver.txt中提供了test網(wǎng)絡(luò)，那么會(huì)在訓(xùn)練完成后進(jìn)行一次測(cè)試

??//?After?the?optimization?is?done,?run?an?additional?train?and?test?pass?to
??//?display?the?train?and?test?loss/outputs?if?appropriate?(based?on?the
??//?display?and?test_interval?settings,?respectively).??Unlike?in?the?rest?of
??//?training,?for?the?train?net?we?only?run?a?forward?pass?as?we've?already
??//?updated?the?parameters?"max_iter"?times?--?this?final?pass?is?only?done?to
??//?display?the?loss,?which?is?computed?in?the?forward?pass.
??if?(param_.display()?&&?iter_?%?param_.display()?==?0)?{
????int?average_loss?=?this->param_.average_loss();
????Dtype?loss;
????net_->Forward(&loss);

????UpdateSmoothedLoss(loss,?start_iter,?average_loss);

????LOG(INFO)?<<?"Iteration?"?<<?iter_?<<?",?loss?=?"?<<?smoothed_loss_;
??}
??if?(param_.test_interval()?&&?iter_?%?param_.test_interval()?==?0)?{
????TestAll();
??}

在Step函數(shù)中通過while循環(huán)迭代訓(xùn)練，并且如果設(shè)置有測(cè)試網(wǎng)絡(luò)，在設(shè)置條件滿足時(shí)，每次循環(huán)會(huì)先對(duì)當(dāng)前網(wǎng)絡(luò)進(jìn)行測(cè)試

??while?(iter_?<?stop_iter)?{
????//?zero-init?the?params
????net_->ClearParamDiffs();
????if?(param_.test_interval()?&&?iter_?%?param_.test_interval()?==?0
????????&&?(iter_?>?0?||?param_.test_initialization()))?{
??????if?(Caffe::root_solver())?{
????????TestAll();
??????}

測(cè)試完成后，如何沒有終止訓(xùn)練，將繼續(xù)訓(xùn)練，此處的iter_size默認(rèn)值是1，主要作用是SGD中參數(shù)更新頻率，即訓(xùn)練iter_size后更新網(wǎng)絡(luò)，此時(shí)訓(xùn)練的總樣本數(shù)為train.txt中定義的batch_size * iter_size。

????for?(int?i?=?0;?i?<?param_.iter_size();?++i)?{
??????loss?+=?net_->ForwardBackward();
????}

之后調(diào)用ApplyUpdate();更新權(quán)值和偏置，更新方法后續(xù)再聊。

Step中的測(cè)試與caffe.cpp中的test類似，主要是檢測(cè)當(dāng)前網(wǎng)絡(luò)訓(xùn)練狀態(tài)，可以根據(jù)任務(wù)狀態(tài)提前終止訓(xùn)練，比如測(cè)試的損失函數(shù)達(dá)到一定范圍。