MOD:multi class label

docs/data_format.md

@@ -12,7 +12,7 @@ Ytk-learn supports weighted trainning and sample weight scaling, so each line is
   - regression:  real number, e.g. 122.23
   - binary classification: 0 or 1
   - binary cross_entropy: real number, belongs to [0, 1], e.g. 0.245
-  - multiclass classficication: one-hot coding,  length of labels is equal to class number, e.g. 0, 0, 1, 0 means that target is class2 in 4 classes(class_0, class_1, class_2, class_3)
+  - multiclass classficication: integer number from 0 to K-1, K is class number.
   - multiclass cross_entropy: length of labels is equal to class number, sum of labels equals to 1, e.g. 0.2,0.1,0.4,0.3 (4 class in total)
 
 
@@ -52,7 +52,11 @@ then the data format becomes
    1###0###height:2.0,weight:50.0,size:80
    ```
 
-   There are two samples, the first '10' in the first line is the sample weight, the '1' in ``10###1###``is the sample label. Similarly, the second sample has weight '1' and label '0'. In binary classification, label '1' stands for positive sample while '0' stands for negative sample. You can also provide probability values in [0,1] as label, indicating the probability that the sample is positive, e.g.
+   There are two samples, the first '10' in the first line is the sample weight, the '1' in ``10###1###``is the sample label. Similarly, the second sample has weight '1' and label '0'. In binary classification, label '1' stands for positive sample while '0' stands for negative sample. 
+
+- **binary cross-entropy**
+
+   You can also provide probability values in [0,1] as label, indicating the probability that the sample is positive, e.g.
 
    ```
    10###0.9###height:1.6,weight:56.0,size:102
@@ -68,10 +72,14 @@ then the data format becomes
 - **multi-class classification**
 
    ```
-   1###0,0,0,1,0,0###height:1.6,weight:56.0,size:102
+   1###3###height:1.6,weight:56.0,size:102
    ```
 
-   The first '1' is the sample weight, '0,0,0,1,0,0' is the sample label which means target is the fourth class of the six classes. Ytklearn also supports probability values in [0,1] as label, to indicate the probability that the sample belongs to this class, e.g.
+   label is 3 means target is the fourth class of the six classes. 
+
+- **multi cross-entropy**
+
+   Ytk-learn also supports probability values in [0,1] as label, to indicate the probability that the sample belongs to this class, e.g.
 
    ```
    1###0.01,0.02,0.01,0.75,0.01,0.2###height:1.6,weight:56.0,size:102

docs/gbdt.config.md

@@ -67,7 +67,7 @@ data {
     #   regression : weight###label###f1name:f1value,f2name:f2value,...(###init_prediction)
     #   binary classification : weight###label(0 or 1)###f1name:f1value,f2name:f2value,...(###init_prediction)
     #   binary cross_entropy : weight###label(0~1, positive)###f1name:f1value,f2name:f2value,...(###init_prediction)
-    #   multi-class classficication : weight###0,0,1,0(total 4 class, only one is 1, others is 0, this belongs to 3'rd class)###f1name:f1value,f2name:f2value,...(###init_prediction)
+    #   multi classficication : weight###2(this belongs to 3'rd class, label must be in range [0,K-1], K is class number)###f1name:f1value,f2name:f2value,...
     #   multi cross_entropy : weight###0.2,0.1,0.4,0.3(total 4 class, sum must be equal 1.0)###f1name:f1value,f2name:f2value,...(###init_prediction)
     # (###init_prediction) is optional. If you provide initial prediction(s) for each sample, set optimization.sample_dependent_base_prediction to true
     delim {

docs/multiclass_linear.config.md

@@ -34,7 +34,7 @@ data {
     },
 
     # delimiters, see data_format.md for more details
-    #   multi classficication : weight###0,0,1,0(total 4 class, only one is 1, others is 0, this belongs to 3'rd class)###f1name:f1value,f2name:f2value,...
+    #   multi classficication : weight###2(this belongs to 3'rd class, label must be in range [0,K-1], K is class number)###f1name:f1value,f2name:f2value,...
     #   multi cross_entropy : weight###0.2,0.1,0.4,0.3(total 4 class, sum must be equal 1.0)###f1name:f1value,f2name:f2value,...
     #   if your model is tree model and 
     delim {

src/main/java/com/fenbi/ytklearn/dataflow/MulticlassLinearModelDataFlow.java

@@ -23,6 +23,7 @@
 
 package com.fenbi.ytklearn.dataflow;
 
+import com.fenbi.ytklearn.exception.YtkLearnException;
 import com.fenbi.ytklearn.feature.FeatureHash;
 import com.fenbi.ytklearn.fs.IFileSystem;
 import com.fenbi.mp4j.exception.Mp4jException;
@@ -103,13 +104,26 @@ protected void updateY() throws Exception {
         protected boolean yExtract(String line, String[] info) throws Exception {
             String []linfo = info[1].split(coreParams.y_delim);
 
-            if (linfo.length != K) {
-                throw new Exception("label num must equal:" + K + ", line:" + line);
+            if (linfo.length != K && linfo.length != 1) {
+                throw new Exception("label num must = " + K + ", or = 1, line:" + line);
             }
-            for (int i = 0; i < K; i++) {
-                label[i] = Float.parseFloat(linfo[i]);
+
+            if (linfo.length == 1) {
+                for (int i = 0; i < K; i++) {
+                    label[i] = 0;
+                }
+                int clazz = Integer.parseInt(linfo[0]);
+                if (clazz >= K) {
+                    throw new YtkLearnException("multi classification label must in range [0,K-1]!\n" + line);
+                }
+                label[clazz] = 1.0f;
+            } else {
+                for (int i = 0; i < K; i++) {
+                    label[i] = Float.parseFloat(linfo[i]);
+                }
             }
 
+
             if (coreParams.needYStat) {
                 labelIdx = -1;
                 for (int i = 0; i < K; i++) {
@@ -125,7 +139,7 @@ protected boolean yExtract(String line, String[] info) throws Exception {
 
 
             if (labelIdx == -1) {
-                throw new Exception("label error! line:" + line);
+                throw new Exception("label error for y sampling! line:" + line);
             }
             float rate = coreParams.ySampling[labelIdx];
             if (rate <= 1.0f) {