func auc(actual []float64, posterior []float64) float64 {
log := util.GetLogger()
r := tied_rank(posterior)
num_positive := 0.
sum_positive := 0.
for i := 0; i < len(actual); i++ {
if actual[i] == 1 {
num_positive++
}
}
num_negative := float64(len(actual)) - num_positive
for i := 0; i < len(r); i++ {
if actual[i] == 1 {
sum_positive += r[i]
}
}
if num_negative*num_positive < 0.00001 {
log.Info(fmt.Sprintf("num_positive %d, num_negative %d, sum_positive%d\n", num_positive, num_negative, sum_positive))
return 0.
}
auc := ((sum_positive - num_positive*(num_positive+1)/2.0) /
(num_negative * num_positive))
return auc
}
func (ft *FtrlTrainer) Initialize(epoch int, cache_feature_num bool) bool {
ft.Epoch = epoch
ft.CacheFeatureNum = cache_feature_num
ft.Init = true
ft.log = util.GetLogger()
return ft.Init
}
func StreamRun(model_file string, instances []string) (string, error) {
log := util.GetLogger()
if !util.FileExists(model_file) || len(instances) == 0 {
log.Error("[Predictor-StreamRun] Model file or instances error.")
return fmt.Sprintf(errorjson, "[Predictor-StreamRun] Model file or instances error."), errors.New("[Predictor-StreamRun] Model file or instances error.")
}
var rtstr string
var model solver.LRModel
model.Initialize(model_file)
for i := 0; i < len(instances); i++ {
res, _, x := util.ParseSample(instances[i])
if res != nil {
break
}
pred := model.Predict(x)
pred = math.Max(math.Min(pred, 1.-10e-15), 10e-15)
if i == len(instances)-1 {
rtstr += strconv.FormatFloat(pred, 'f', 6, 64)
} else {
rtstr += strconv.FormatFloat(pred, 'f', 6, 64) + ","
}
}
return fmt.Sprintf(streamjson, rtstr), nil
}
func (lr *LRModel) Initialize(path string) error {
lr.log = util.GetLogger()
file, err := os.Open(path)
if err != nil {
lr.log.Error(fmt.Sprintf("[LRModel-Initialize] Lr model initialize error.%s", err.Error()))
return errors.New(fmt.Sprintf("[LRModel-Initialize] Lr model initialize error.%s", err.Error()))
}
defer file.Close()
var fls FtrlSolver
m, err := ioutil.ReadAll(file)
if err != nil {
lr.log.Error(fmt.Sprintf("[LRModel-Initialize] Lr model initialize error.%s", err.Error()))
return errors.New(fmt.Sprintf("[LRModel-Initialize] Lr model initialize error.%s", err.Error()))
}
err2 := json.Unmarshal(m, &fls)
if err2 != nil {
lr.log.Error(fmt.Sprintf("[LRModel-Initialize] Lr model initialize error.%s", err2.Error()))
return errors.New(fmt.Sprintf("[LRModel-Initialize] Lr model initialize error.%s", err2.Error()))
}
lr.Model = make(map[int]float64)
for i := 0; i < len(fls.Weights); i++ {
lr.Model[fls.Weights[i].Index] = fls.Weights[i].Value
}
lr.Init = true
return nil
}
func (fs *FtrlSolver) SaveModel(path string) error {
log := util.GetLogger()
if !fs.Init {
log.Error("[FtrlSolver-SaveModel] Ftrl solver initialize error.")
return errors.New("[FtrlSolver-SaveModel] Ftrl solver initialize error.")
}
file, err := os.Create(path)
if err != nil {
log.Error(fmt.Sprintf("[FtrlSolver-SaveModel] Ftrl solver save model error.%s", err.Error()))
return errors.New(fmt.Sprintf("[FtrlSolver-SaveModel] Ftrl solver save model error.%s", err.Error()))
}
fs.Weights = make(util.Pvector, fs.Featnum)
for i := 0; i < fs.Featnum; i++ {
val := util.Round(fs.GetWeight(i), 5)
fs.Weights[i] = util.Pair{i, val}
}
b, err2 := json.Marshal(fs)
if err2 != nil {
log.Error(fmt.Sprintf("[FtrlSolver-SaveModel] Ftrl solver save model error.%s", err2.Error()))
return errors.New(fmt.Sprintf("[FtrlSolver-SaveModel] Ftrl solver save model error.%s", err2.Error()))
}
_, err = file.Write(b)
if err != nil {
log.Error(fmt.Sprintf("[FtrlSolver-SaveModel] Ftrl solver save model error.%s", err2.Error()))
return errors.New(fmt.Sprintf("[FtrlSolver-SaveModel] Ftrl solver save model error.%s", err2.Error()))
}
return nil
}
func (lan *Lands) Initialize(configFile string) error {
var err error
lan.mux = make(map[string]func(http.ResponseWriter, *util.ModelParam) error)
lan.mux["/goline/online"] = lan.onlineServeHttp
lan.mux["/goline/offline"] = lan.offlineServeHttp
lan.mux["/goline/predict"] = lan.predictServeHttp
file, err := ioutil.ReadFile(configFile)
if err != nil {
lan.log4goline.Error("[Lands-Initialize]Open Config error." + err.Error())
return errors.New("[Lands-Initialize]Open Config error." + err.Error())
}
temp := new(util.Config)
if err = json.Unmarshal(file, temp); err != nil {
lan.log4goline.Error("[Lands-Initialize]Parse config file error." + err.Error())
return errors.New("[Lands-Initialize]Parse config file error." + err.Error())
}
lan.conf = temp
lan.pool = util.InitRedisPool(&lan.conf.Redis)
if lan.pool == nil {
lan.log4goline.Error("[Lands-Initialize]Initialize redis pool error.")
return errors.New("[Lands-Initialize]Initialize redis pool error.")
}
util.InitLogger(lan.conf.LogModule)
lan.log4goline = util.GetLogger()
return nil
}
func (lft *LockFreeFtrlTrainer) Initialize(
epoch int,
num_threads int,
cache_feature_num bool) bool {
lft.Epoch = epoch
lft.CacheFeatureNum = cache_feature_num
lft.NumThreads = num_threads
lft.log = util.GetLogger()
lft.Init = true
return lft.Init
}
func (fps *FtrlParamServer) Initialize(
alpha float64,
beta float64,
l1 float64,
l2 float64,
n int,
dropout float64) error {
fps.log = util.GetLogger()
if !fps.FtrlSolver.Initialize(alpha, beta, l1, l2, n, dropout) {
fps.log.Error("[FtrlParamServer-Initialize] Fast ftrl solver initialize error.")
return errors.New("[FtrlParamServer-Initialize] Fast ftrl solver initialize error.")
}
fps.ParamGroupNum = calc_group_num(n)
fps.LockSlots = make([]sync.Mutex, fps.ParamGroupNum)
fps.Init = true
return nil
}
func (fs *FtrlSolver) SaveEncodeModel() (string, error) {
log := util.GetLogger()
if !fs.Init {
log.Error("[FtrlSolver-SaveEncodeModel] Ftrl solver initialize error.")
return "", errors.New("[FtrlSolver-SaveEncodeModel] Ftrl solver initialize error.")
}
fs.Weights = make(util.Pvector, fs.Featnum)
for i := 0; i < fs.Featnum; i++ {
val := util.Round(fs.GetWeight(i), 5)
fs.Weights[i] = util.Pair{i, val}
}
b, err := json.Marshal(fs)
if err != nil {
log.Error(fmt.Sprintf("[FtrlSolver-SaveEncodeModel] Ftrl solver save model error.%s", err.Error()))
return "", errors.New(fmt.Sprintf("[FtrlSolver-SaveEncodeModel] Ftrl solver save model error.%s", err.Error()))
}
return string(b), nil
}
func (fft *FastFtrlTrainer) Initialize(
epoch int,
num_threads int,
cache_feature_num bool,
burn_in float64,
push_step int,
fetch_step int) bool {
fft.Epoch = epoch
fft.CacheFeatureNum = cache_feature_num
fft.PusStep = push_step
fft.FetchStep = fetch_step
if num_threads == 0 {
fft.NumThreads = runtime.NumCPU()
} else {
fft.NumThreads = num_threads
}
fft.Init = true
fft.BurnIn = burn_in
fft.log4fft = util.GetLogger()
return fft.Init
}
func (fw *FtrlWorker) Initialize(
param_server *FtrlParamServer,
push_step int,
fetch_step int) bool {
fw.FtrlSolver.Alpha = param_server.Alpha
fw.FtrlSolver.Beta = param_server.Beta
fw.FtrlSolver.L1 = param_server.L1
fw.FtrlSolver.L2 = param_server.L2
fw.FtrlSolver.Featnum = param_server.Featnum
fw.FtrlSolver.Dropout = param_server.Dropout
fw.NUpdate = make([]float64, fw.FtrlSolver.Featnum)
fw.ZUpdate = make([]float64, fw.FtrlSolver.Featnum)
fw.SetFloatZero(fw.NUpdate, fw.FtrlSolver.Featnum)
fw.SetFloatZero(fw.ZUpdate, fw.FtrlSolver.Featnum)
fw.N = make([]float64, fw.FtrlSolver.Featnum)
fw.Z = make([]float64, fw.FtrlSolver.Featnum)
if param_server.FetchParam(fw.N, fw.Z) != nil {
return false
}
fw.ParamGroupNum = calc_group_num(fw.FtrlSolver.Featnum)
fw.ParamGroupStep = make([]int, fw.ParamGroupNum)
for i := 0; i < fw.ParamGroupNum; i++ {
fw.ParamGroupStep[i] = 0
}
fw.PushStep = push_step
fw.FetchStep = fetch_step
fw.log = util.GetLogger()
fw.FtrlSolver.Init = true
return fw.FtrlSolver.Init
}
func print_usage(argc int, argv []string) {
log := util.GetLogger()
log.Error(fmt.Sprintf("Usage:\n", ""))
log.Error(fmt.Sprintf("\t%s job_name test_file model output_file threshold\n", argv[0]))
}
func Run(argc int, argv []string) (string, error) {
var job_name string
var test_file string
var model_file string
var output_file string
var threshold float64
log := util.GetLogger()
if len(argv) == 5 {
job_name = argv[0]
test_file = argv[1]
model_file = argv[2]
output_file = argv[3]
threshold, _ = strconv.ParseFloat(argv[4], 64)
} else {
print_usage(argc, argv)
log.Error("[Predictor-Run] Input parameters error.")
return fmt.Sprintf(errorjson, "[Predictor-Run] Input parameters error."), errors.New("[Predictor-Run] Input parameters error.")
}
if len(job_name) == 0 || len(test_file) == 0 || len(model_file) == 0 || len(output_file) == 0 {
print_usage(argc, argv)
log.Error("[Predictor-Run] Input parameters error.")
return fmt.Sprintf(errorjson, "[Predictor-Run] Input parameters error."), errors.New("[Predictor-Run] Input parameters error.")
}
var model solver.LRModel
model.Initialize(model_file)
var wfp *os.File
var err1 error
exist := func(filename string) bool {
var exist = true
if _, err := os.Stat(filename); os.IsNotExist(err) {
exist = false
}
return exist
}
if exist(output_file) {
wfp, err1 = os.OpenFile(output_file, os.O_SYNC, 0666)
} else {
wfp, err1 = os.Create(output_file)
}
if err1 != nil {
log.Error("[Predictor-Run] Open file error." + err1.Error())
return fmt.Sprintf(errorjson, err1.Error()), errors.New("[Predictor-Run] Open file error." + err1.Error())
}
defer wfp.Close()
cnt := 0 //样本总数
pcorrect := 0 //正样本预测正确数
pcnt := 0 //正样本总数
ncorrect := 0 //负样本预测正确数
var loss float64 = 0.
var parser trainer.FileParser
err := parser.OpenFile(test_file)
if err != nil {
log.Error("[Predictor-Run] Open file error." + err.Error())
return fmt.Sprintf(errorjson, err.Error()), errors.New("[Predictor-Run] Open file error." + err.Error())
}
var pred_scores util.Dvector
for {
res, y, x := parser.ReadSample()
if res != nil {
break
}
pred := model.Predict(x)
pred = math.Max(math.Min(pred, 1.-10e-15), 10e-15)
wfp.WriteString(fmt.Sprintf("%f\n", pred))
pred_scores = append(pred_scores, util.DPair{pred, y})
cnt++
if util.UtilFloat64Equal(y, 1.0) {
pcnt++
}
var pred_label float64 = 0
if pred > threshold {
pred_label = 1
}
if util.UtilFloat64Equal(pred_label, y) {
if util.UtilFloat64Equal(y, 1.0) {
pcorrect++
} else {
ncorrect++
}
}
pred = math.Max(math.Min(pred, 1.-10e-15), 10e-15)
if y > 0 {
loss += -math.Log(pred)
} else {
loss += -math.Log(1. - pred)
}
}
auc := calc_auc(pred_scores)
if auc < 0.5 {
auc = 0.5
}
if cnt > 0 {
log.Info(fmt.Sprintf("[%s] Log-likelihood = %f\n", job_name, float64(loss)/float64(cnt)))
log.Info(fmt.Sprintf("[%s] Precision = %.2f%% (%d/%d)\n", job_name,
float64(pcorrect*100)/float64(cnt-pcnt-ncorrect+pcorrect),
pcorrect, cnt-pcnt-ncorrect+pcorrect))
log.Info(fmt.Sprintf("[%s] Recall = %.2f%% (%d/%d)\n", job_name,
float64(pcorrect*100)/float64(pcnt), pcorrect, pcnt))
log.Info(fmt.Sprintf("[%s] Accuracy = %.2f%% (%d/%d)\n", job_name,
float64((pcorrect+ncorrect)*100)/float64(cnt), (pcorrect + ncorrect), cnt))
log.Info(fmt.Sprintf("[%s] AUC = %f\n", job_name, auc))
}
parser.CloseFile()
util.Write2File(output_file, fmt.Sprintf(" Log-likelihood = %f\n Precision = %f (%d/%d)\n Recall = %f (%d/%d)\n Accuracy = %f (%d/%d)\n AUC = %f\n",
float64(loss)/float64(cnt),
float64(pcorrect)/float64(cnt-pcnt-ncorrect+pcorrect), pcorrect, cnt-pcnt-ncorrect+pcorrect,
float64(pcorrect)/float64(pcnt), pcorrect, pcnt,
float64(pcorrect+ncorrect)/float64(cnt), pcorrect+ncorrect, cnt,
auc))
return fmt.Sprintf(returnJson,
job_name,
fmt.Sprintf("Log-likelihood = %f", float64(loss)/float64(cnt)),
fmt.Sprintf("Precision = %f (%d/%d)", float64(pcorrect)/float64(cnt-pcnt-ncorrect+pcorrect), pcorrect, cnt-pcnt-ncorrect+pcorrect),
fmt.Sprintf("Recall = %f (%d/%d)", float64(pcorrect)/float64(pcnt), pcorrect, pcnt),
fmt.Sprintf("Accuracy = %f (%d/%d)", float64((pcorrect+ncorrect))/float64(cnt), (pcorrect+ncorrect), cnt),
fmt.Sprintf("AUC = %f", auc),
output_file), nil
}
func read_problem_info(
train_file string,
read_cache bool,
num_threads int) (int, int, error) {
feat_num := 0
line_cnt := 0
log := util.GetLogger()
var lock sync.Mutex
var parser FileParser
var errall error
read_from_cache := func(path string) error {
fs, err := os.Open(path)
defer fs.Close()
if err != nil {
return err
}
bfRd := bufio.NewReader(fs)
line, err := bfRd.ReadString('\n')
if err != nil {
return err
}
var res []string = s.Split(line, " ")
if len(res) != 2 {
log.Error("[read_problem_info] File format error.")
return errors.New("[read_problem_info] File format error.")
}
feat_num, errall = strconv.Atoi(res[0])
if errall != nil {
log.Error("[read_problem_info] Label format error." + errall.Error())
return errors.New("[read_problem_info] Label format error." + errall.Error())
}
line_cnt, errall = strconv.Atoi(res[1])
if errall != nil {
log.Error("[read_problem_info] Feature format error." + errall.Error())
return errors.New("[read_problem_info] Feature format error." + errall.Error())
}
return nil
}
exist := func(filename string) bool {
var exist = true
if _, err := os.Stat(filename); os.IsNotExist(err) {
exist = false
}
return exist
}
write_to_cache := func(filename string) error {
var f *os.File
var err1 error
if exist(filename) {
f, err1 = os.OpenFile(filename, os.O_WRONLY, 0666)
} else {
f, err1 = os.Create(filename)
}
if err1 != nil {
return err1
}
defer f.Close()
wireteString := string(feat_num) + " " + string(line_cnt) + "\n"
_, err1 = io.WriteString(f, wireteString)
if err1 != nil {
return err1
}
return nil
}
read_problem_worker := func(i int, c *sync.WaitGroup) {
local_max_feat := 0
local_count := 0
for {
flag, _, local_x := parser.ReadSampleMultiThread()
if flag != nil {
break
}
for i := 0; i < len(local_x); i++ {
if local_x[i].Index+1 > local_max_feat {
local_max_feat = local_x[i].Index + 1
}
}
local_count++
}
lock.Lock()
line_cnt += local_count
lock.Unlock()
if local_max_feat > feat_num {
feat_num = local_max_feat
}
defer c.Done()
}
cache_file := string(train_file) + ".cache"
cache_exists := exist(cache_file)
if read_cache && cache_exists {
read_from_cache(cache_file)
} else {
parser.OpenFile(train_file)
util.UtilParallelRun(read_problem_worker, num_threads)
parser.CloseFile()
}
log.Info(fmt.Sprintf("[read_problem_info] Instances=[%d] features=[%d]\n", line_cnt, feat_num))
if read_cache && !cache_exists {
write_to_cache(cache_file)
}
return feat_num, line_cnt, nil
}
