How Do I Get A Proper Prediction With Libsvm?

Question

[acampbe2]

I've just recently started using libsvm. I've been trying for awhile to train my features get the model and then do a prediction with corresponding features. Unforunately I contiously get wrong file format. I've routinely did this measure cut my features and then train on them, which i get a model.

This is how I train to get the model.

./svm-train -t 0 -b 1 svmtrainModel1Revised.txt TrainedModel1.svm

0 1:8.109121    2:11.063075 3:9.265099 4:7.392232  5:9.577386
0 1:7.497237    2:10.314889 3:9.330452 4:8.658748 5:9.493017
0 1:7.579283    2:9.853208 3:9.863712 4:8.48306 5:9.84812
1 1:7.683603    2:10.323639 3:10.616027 4:7.039438 5:10.321418
1 1:7.645213    2:9.742212 3:9.957533 4:8.810831 5:8.86067
0 1:7.738999    2:9.956453 3:9.643299 4:8.553764 5:9.824224
1 1:7.968907    2:10.610946 3:9.061123 4:7.84069 5:9.793263

The model I get from above I predict based on that.

./svm-predict -b 1 svmtest1RevisedA.txt TrainedModel1.svm labelmodel1.txt

These features are in the same format as above, except they are lacking the ground truth.

8.139867      10.146595    9.554127    8.409128    9.380923
7.691873    9.820682    9.777709    9.136581    9.038513
8.629565    10.684548    9.192018    9.031779    9.733009
7.796154    10.871736    9.113837    7.329961    10.071653
7.235544    9.994834    10.601161    8.720384    10.112152
8.008006    10.75276    10.755063    7.517268    9.415618
7.857692    9.936756    9.556366    8.941176    8.933603
7.806821    9.934397    9.324827    8.422852    8.797105

There are no commas and there is only a space between each element,but when i got to predict it tells me "Wrong input format at line 1" please help.

Answer 1

[Michael Schubert]

I haven't used libsvm much, but your models seems to be missing the header. This is what I get when I run the svm-train command (same software version, package for ArchLinux):

svm_type c_svc
kernel_type linear
nr_class 2
total_sv 6
rho 2.76767
label 0 1
probA 1.12362
probB -1.22234
nr_sv 3 3
SV
1 1:8.109121 2:11.063075 3:9.265099 4:7.392232 5:9.577386
1 1:7.579283 2:9.853208 3:9.863712 4:8.48306 5:9.84812
1 1:7.738999 2:9.956453 3:9.643299 4:8.553764 5:9.824224
-1 1:7.683603 2:10.323639 3:10.616027 4:7.039438 5:10.321418
-1 1:7.645213 2:9.742212 3:9.957533 4:8.810831 5:8.86067
-1 1:7.968907 2:10.610946 3:9.061123 4:7.84069 5:9.793263

Make sure that your file didn't get corrupted andMake sure that your file didn't get corrupted and you don't get any errors when training your data. Another reason might be that the libsvm package of your distribution is broken.

Answer 2

[Chris]

Your set containing the instances to be predicted needs to be in exactly the same format as the training data. This means it needs a leading (dummy) class label (set it to 1) per instance, and the feature values per instance have to be preceded by a running number that reflects the feature number plus colon.