1、背景html
本來是打算按《DEX Deep EXpectation of apparent age from a single image》進行表面年齡的訓練,可因爲IMDB-WIKI的數據集比較龐大,各個年齡段分佈不均勻,難以劃分訓練集及驗證集。後來爲了先跑通整個訓練過程的主要部分,就直接用LAP數據集,參考caffe的finetune_flickr_style,進行一些參數修改,利用bvlc_reference_caffenet.caffemodel完成年齡估計的finetune。ios
2、訓練數據集準備git
一、下載LAP數據集,包括Train、Validation、Test,以及對應的年齡label,http://chalearnlap.cvc.uab.es/dataset/18/description/,須要註冊。github
二、將標註好的csv文件轉換爲caffe識別的txt格式。csv每一行的信息爲:圖片名、年齡、標準差。訓練的時候不須要標準差信息,咱們只要將圖片名和年齡寫入到txt中,並按空格隔開,獲得Train.txt以下:app
一樣,完成驗證集cvs文件的轉換,獲得Validation.txt。函數
3、模型及相關文件拷貝工具
一、拷貝預訓練好的vgg16模型caffe\models\bvlc_reference_caffenet\bvlc_reference_caffenet.caffemodel至工做目錄下,該文件約232M;測試
二、拷貝caffe\models\finetune_flickr_style文件夾中deploy.prototxt、solver.prototxt、train_val.prototxt至工做目錄下;spa
三、拷貝imageNet的均值文件caffe\data\ilsvrc12\imagenet_mean.binaryproto至工做目錄下。code
4、參數修改
一、修改train_val.prototxt
以及最後的輸出層個數,由於咱們要訓練的爲[0,100]歲的輸出,共101類,因此:
二、修改solver.protxt
三、修改用於實際測試的部署文件deploy.protxt
輸出層的個數也要改:
5、開始訓練
一、新建train.bat
caffe train -solver solver.prototxt -weights bvlc_reference_caffenet.caffemodel rem caffe train --solver solver.prototxt --snapshot snapshot/bvlc_iter_48000.solverstate pause
雙擊便可開始訓練,當訓練過程當中出現意外中斷,可註釋第一行,關閉第二行註釋,根據實際狀況修改保存,繼續雙擊訓練。
個人電腦CPU是i5 6500,顯卡爲gtx1050Ti,8G內存,大體要訓練10個小時吧,中途也出現了一些內存不足訓練終止的狀況。
二、訓練結束
6、模型評價
年齡估計本來是一個線性問題,不是一個明確的分類問題,人都沒法準確無誤地獲得某人的年齡,更況且是機器呢。因此評價這個年齡分類模型的好壞不能簡單地經過精度來衡量,能夠用MAE(平均絕對偏差)以及ε-error來衡量,其中
一、對驗證集Validation.txt的全部圖片進行預測
藉助 https://github.com/eveningglow/age-and-gender-classification ,其環境搭建可參考http://www.javashuo.com/article/p-tkoovjxg-o.html
修改main函數
int split(std::string str, std::string pattern, std::vector<std::string> &words) { std::string::size_type pos; std::string word; int num = 0; str += pattern; std::string::size_type size = str.size(); for (auto i = 0; i < size; i++) { pos = str.find(pattern, i); if (pos == i) { continue;//if first string is pattern } if (pos < size) { word = str.substr(i, pos - i); words.push_back(word); i = pos + pattern.size() - 1; num++; } } return num; } //param example: model/deploy_age2.prototxt model/age_net.caffemodel model/mean.binaryproto img/0008.jpg int main(int argc, char* argv[]) { if (argc != 5) { cout << "Command shoud be like ..." << endl; cout << "AgeAndGenderClassification "; cout << " \"AGE_NET_MODEL_FILE_PATH\" \"AGE_NET_WEIGHT_FILE_PATH\" \"MEAN_FILE_PATH\" \"TEST_IMAGE\" " << endl; std::cout << "argc = " << argc << std::endl; getchar(); return 0; } // Get each file path string age_model(argv[1]); string age_weight(argv[2]); string mean_file(argv[3]); //string test_image(argv[4]); // Probability vector vector<Dtype> prob_age_vec; // Set mode Caffe::set_mode(Caffe::GPU); // Make AgeNet AgeNet age_net(age_model, age_weight, mean_file); // Initiailize both nets age_net.initNetwork(); //讀取待測試的圖片名 std::ifstream fin("E:\\caffe\\DEX_age_gender_predict\\lap2\\Validation.txt"); std::string line; std::vector<std::string> test_images; std::vector<int> test_images_age; while (!fin.eof()) { std::getline(fin, line); std::vector<std::string> words; split(line, " ", words); test_images.push_back(words[0]); test_images_age.push_back(atoi(words[1].c_str())); } std::cout << "test_images size = " << test_images.size() << std::endl; std::ofstream fout("E:\\caffe\\DEX_age_gender_predict\\lap2\\Validation_predict.txt"); for (int k = 0; k < test_images.size(); ++k) { std::cout << "k = " << k << std::endl; std::string test_image; test_image = test_images[k]; // Classify and get probabilities Mat test_img = imread(test_image, CV_LOAD_IMAGE_COLOR); int age = age_net.classify(test_img, prob_age_vec); // Print result and show image //std::cout << "prob_age_vec size = " << prob_age_vec.size() << std::endl; //for (int i = 0; i < prob_age_vec.size(); ++i) { // std::cout << "[" << i << "] = " << prob_age_vec[i] << std::endl; //} //Dtype prob; //int index; //get_max_value(prob_age_vec, prob, index); //std::cout << "prob = " << prob << ", index = " << index << std::endl; //imshow("AgeAndGender", test_img); //waitKey(0); fout << test_images[k] << " " << test_images_age[k] << " " << age << std::endl; } std::cout << "finish!" << std::endl; getchar(); return 0; }
個人命令參數爲:E:\caffe\DEX_age_gender_predict\lap2\deploy.prototxt E:\caffe\DEX_age_gender_predict\lap2\snapshot\bvlc_iter_50000.caffemodel model\mean.binaryproto img\0008.jpg
可根據實際狀況修改。可獲得Validation_predict.txt文件。運行過程當中可能會由於內存不足中斷運行,可能要分批次運行屢次。
二、計算MAE及ε-error
(1)將Validation_predict.txt文件及驗證集的標註文件Reference.csv拷貝到新建的vs項目的工做目錄下;
(2)計算
#include <iostream> #include <string> #include <fstream> #include <vector> int split(std::string str, std::string pattern, std::vector<std::string> &words) { std::string::size_type pos; std::string word; int num = 0; str += pattern; std::string::size_type size = str.size(); for (auto i = 0; i < size; i++) { pos = str.find(pattern, i); if (pos == i) { continue;//if first string is pattern } if (pos < size) { word = str.substr(i, pos - i); words.push_back(word); i = pos + pattern.size() - 1; num++; } } return num; } int main(int argc, char** argv) { //u, sigma, x std::vector<int> u; std::vector<float> sigma; std::vector<int> predict; std::string line; std::ifstream csv_file("Reference.csv"); while (!csv_file.eof()) { std::getline(csv_file, line); std::vector<std::string> words; split(line, ";", words); u.push_back(atoi(words[1].c_str())); sigma.push_back(atof(words[2].c_str())); } std::ifstream predict_file("Validation_predict.txt"); while (!predict_file.eof()) { std::getline(predict_file, line); std::vector<std::string> words; split(line, " ", words); predict.push_back(atoi(words[2].c_str())); } if (u.size() != predict.size()) { std::cout << "u.size() != predict.size()" << std::endl; getchar(); return -1; } //MAE int sum_err = 0; float MAE = 0; for (int i = 0; i < u.size(); ++i) { sum_err += abs(u[i] - predict[i]); } MAE = static_cast<float>(sum_err) / u.size(); std::cout << "MAE = " << MAE << std::endl;//11.7184 //esro-error std::vector<float> errors; float err = 0; float error = 0.0; for (int i = 0; i < u.size(); ++i) { err = 1.0 - exp(-1.0*(predict[i] - u[i])*(predict[i] - u[i]) / (2 * sigma[i] * sigma[i])); errors.push_back(err); error += err; } error /= errors.size(); std::cout << "error = " << error << std::endl;//0.682652 std::cout << "finish!" << std::endl; getchar(); return 0; }
最終獲得MAE爲11.7184, ε-error爲0.682652。
7、實際應用中預測
一、可利用caffe提供的classification工具對輸入圖片地進行估計
classification deploy.prototxt snapshot\bvlc_iter_50000.caffemodel imagenet_mean.binaryproto ..\age_labels.txt ..\test_image\test_3.jpg
pause
其中,age_labels.txt爲0-100個label的說明信息,每一個label對應一行,共101行,個人寫法以下:
end