LIBSVM: output of svm-predict is all 0 even though I'm using the test data which was used for training

Question

I'm using libsvm to perform binary classification. I used easy.py for training which is included in libsvm library. After running easy.py, it showed the following in stdout.

Cross validation...
Best c=8192.0, g=8.0 CV rate=93.3155
Training...
Output model: parameters.data.model

The svm model file is generated, and the accuracy looks good. So I used svm-predict (executable file) to see if the generated SVM really classifies data. Here, the test data I used is exactly the same as the one I used for the training.

However all test data is classified as label 0. and the accuracy shown is really low. The command I executed is something like following.

python easy.py parameters.data
# model and scale files are generated

./svm-predict parameters.data parameters.data.model svmpredict_output.log
# svmpredict_output.log is all 0

Am I doing something wrong?

Answer 1

Ok, sorry. My stupid mistake.

The svm used scaled dataset for training, but I was using un-scaled dataset for test data.

./svm-predict parameters.data.scale parameters.data.model svmpredict_output.log

Like above, everything works well. (extension like *.scale, *.data should be the same if you're using easy.py)