Multi-labels using two different LMDB

Question

I am new in caffe framework and I would like to use caffe to implement the training with multi-label. I use two LMDB to save data and labels, respectively. The data LMDB is of dimension Nx1xHxW while the label LMDB is of dimension Nx1x1x3. Labels are float data.

The text file is as follow:

5911 3
train/train_data/4224.bmp        13         0        12
train/train_data/3625.bmp        11         3         7
...                              ...

I use C++ to create LMDB. My main.cpp:

#include <algorithm>
#include <fstream>  // NOLINT(readability/streams)
#include <string>
#include <utility>
#include <vector>
#include <QImage>

#include "boost/scoped_ptr.hpp"
#include "gflags/gflags.h"
#include "glog/logging.h"

#include "caffe/proto/caffe.pb.h"
#include "caffe/util/db.hpp"
#include "caffe/util/format.hpp"
#include "caffe/util/rng.hpp"

#include <boost/filesystem.hpp>
#include <iomanip>
#include <iostream>  // NOLINT(readability/streams)
#include <string>

#include "google/protobuf/message.h"

#include "caffe/common.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/util/format.hpp"

#ifndef CAFFE_TMP_DIR_RETRIES
#define CAFFE_TMP_DIR_RETRIES 100
#endif

using namespace caffe;  // NOLINT(build/namespaces)
using std::pair;
using boost::scoped_ptr;

const char *dat_lab="/home/mul/caffe-master/examples/2D_3D/new/info/train.data";
string data_db="/home/mul/caffe-master/examples/2D_3D/new/2D_3D_data_leveldb";
string label_db="/home/mul/caffe-master/examples/2D_3D/new/2D_3D_label_leveldb";
string root="/home/mul/caffe-master/examples/2D_3D/new/";
string path;

int main()
{

    //Create data DB
    scoped_ptr<db::DB> dat_db(db::GetDB("leveldb"));
    dat_db->Open(data_db, db::NEW);
    scoped_ptr<db::Transaction> dat_txn(dat_db->NewTransaction());

    //Create label DB
    scoped_ptr<db::DB> lab_db(db::GetDB("leveldb"));
    lab_db->Open(label_db, db::NEW);
    scoped_ptr<db::Transaction> lab_txn(lab_db->NewTransaction());

    //Storing to db
    Datum dat_datum,lab_datum;
    int count=0;

    std::ifstream infile(dat_lab);
    std::string filename;
    const char *dataname;
    int dataNum;
    int labelcount;
    QImage img;
    infile>>dataNum>>labelcount;
    LOG(INFO) << "A total of " << dataNum<< " images.";

    for (int line_id = 0; line_id < dataNum; ++line_id)
    {
        infile>>filename;
        path=root+filename;
        dataname=path.c_str();
        img.load(dataname);

        dat_datum.set_channels(1);
        dat_datum.set_height(img.height());
        dat_datum.set_width(img.width());
        dat_datum.clear_data();
        dat_datum.clear_float_data();

        int datum_channels = dat_datum.channels();
        int datum_height = dat_datum.height();
        int datum_width = dat_datum.width();
        int datum_size = datum_channels * datum_height * datum_width;
        std::string buffer(datum_size, ' ');
        const uchar* ptr = img.bits();
        int img_index = 0;
        for (int h = 0; h < datum_height; ++h)
        {

            for (int w = 0; w < datum_width; ++w)
            {
                for (int c = 0; c < datum_channels; ++c)
                {
                    int datum_index = (c * datum_height + h) * datum_width + w;
                    buffer[datum_index] = static_cast<char>(ptr[img_index++]);
                }
            }
        }
        dat_datum.set_data(buffer);


        lab_datum.set_channels(labelcount);
        lab_datum.set_height(1);
        lab_datum.set_width(1);
        lab_datum.clear_data();
        lab_datum.clear_float_data();
        for(int i=0;i<labelcount;++i)
        {
            float mid;
            infile>>mid;
            lab_datum.add_float_data(mid);
        }

        // sequential
        string key_str = caffe::format_int(line_id, 8);

        // Put in db
        string out;
        CHECK(dat_datum.SerializeToString(&out));
        dat_txn->Put(key_str, out);
        CHECK(lab_datum.SerializeToString(&out));
        lab_txn->Put(key_str, out);

        if (++count % 1000 == 0)
        {
            // Commit db
            dat_txn->Commit();
            dat_txn.reset(dat_db->NewTransaction());
            lab_txn->Commit();
            lab_txn.reset(lab_db->NewTransaction());
            LOG(INFO) << "Processed " << count << " files.";
        }
    }
    // write the last batch
    if (count % 1000 != 0)
    {
        dat_txn->Commit();
        lab_txn->Commit();
        LOG(INFO) << "Processed " << count << " files.";
    }
    return 0;
}

Two LMDB can be created successfully. But when I use caffe to implement the training with two LMDB, the result is always wrong. The loss layer is EUCLIDEAN_LOSS and the loss can not descent. I don't know whether the code which can create two LMDB is wrong. Who can help me ? Thanks whatever.

Answer 1

As a whole, your code above is Ok but you should notice that:

1. Your .cpp is to create LEVELDB instead of LMDB, of course this is not the reason that caused your problem, either type is ok.
2. The "label LMDB" generated by your code is of dimension Nx3x1x1 instead of Nx1x1x3(NumberxChannelxWidthxHeight).
3. In a learning task using minibatch SGD, as far as I know, it is very useful to shuffle your data for training to get more optimized model. I'm not sure whether you noticed this. But at least your cpp didn't shuffle your "train.data".
4. Most important of all, the reason that caused your problem here most possibly lied in the data layer in your network which read your data and label lmdb/leveldb files, because you assigned the labels to the float data of datum and the DataLayer in caffe in fact does not read the float data(Only if you used your self-defined data layer). So please also upload your prototxt file which defines your network. Thus we can find out what the problem really was.

At last, I added a "MultiTaskData" layer MultiTaskDataLayer to read multi-labels from datum for multi-task training and you can make simple modifications to add it to your caffe and use like this:

    name: "AgeNet"
    layer {
        name: "Age"
        type: "MultiTaskData"
        top: "data"
        top: "age_label"
        top: "gender_label"
        data_param { 
            source: "age_gender_classification_0_60p_train_leveldb"   
            batch_size: 60 
            task_num: 2
            label_dimension: 1
            label_dimension: 1
        }
        transform_param {
            scale: 0.00390625
            crop_size: 60
            mirror: true
        }
        include:{ phase: TRAIN }
    }
    layer { 
        name: "cls_age" 
        type: "InnerProduct"
        bottom: "data"  
        top: "cls_age" 
        param {
            lr_mult: 1
            decay_mult: 1
        }
        param {
            lr_mult: 2
            decay_mult: 0
        }
        inner_product_param {
            num_output: 7
            weight_filler {
            type: "xavier"
            }    
        bias_filler {      
            type: "constant"
            }  
        }
    }
    layer {  
        name: "age_loss"  
        type: "SoftmaxWithLoss"  
        bottom: "cls_age" 
        bottom: "age_label"
        top: "age_loss"
        include:{ phase: TRAIN }
    }
    layer { 
        name: "cls_gender" 
        type: "InnerProduct"
        bottom: "data"  
        top: "cls_gender" 
        param {
            lr_mult: 1
            decay_mult: 1
        }
        param {
            lr_mult: 2
            decay_mult: 0
        }
        inner_product_param {
            num_output: 2
            weight_filler {
                type: "xavier"
            }    
            bias_filler {      
                type: "constant"
            }  
        }
    }
    layer {  
        name: "gender_loss"  
        type: "SoftmaxWithLoss"  
        bottom: "cls_gender" 
        bottom: "gender_label"
        top: "gender_loss"
        include:{ phase: TRAIN }
    }