Regression demo is broken

[hcho3]

The demo demo/regression is not working.

$ python mapfeat.py
$ python mknfold.py machine.txt 1
$ ../../xgboost machine.conf
[19:27:05] 175x36 matrix with 1225 entries loaded from machine.txt.train
[19:27:05] 34x34 matrix with 238 entries loaded from machine.txt.test
terminate called after throwing an instance of 'dmlc::Error'
  what():  [19:27:05] /home/phcho/Desktop/xgboost/src/learner.cc:1062: Check failed: 
  learner_model_param_.num_feature == p_fmat->Info().num_col_ (36 vs. 34) : 
  Number of columns does not match number of features in booster.
Stack trace:
  [bt] (0) ../../xgboost(+0x2b8bc) [0x561330f978bc]
  [bt] (1) ../../xgboost(+0x133546) [0x56133109f546]
  [bt] (2) ../../xgboost(+0x138f3a) [0x5613310a4f3a]
  [bt] (3) ../../xgboost(+0x253aa) [0x561330f913aa]
  [bt] (4) ../../xgboost(+0x27e58) [0x561330f93e58]
  [bt] (5) /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7f0ee3a94b97]
  [bt] (6) ../../xgboost(+0x238ba) [0x561330f8f8ba]


Aborted (core dumped)

[hcho3]

When loading from LIBSVM, we should be able to append empty columns to run predictions, because the test data may not necessarily have all the features. This is because the LIBSVM format does not specify the number of columns in the data.

[trivialfis]

I prefer leaving it to data preprocessing.

[hcho3]

I prefer leaving it to data preprocessing.

Not sure what you mean by this. The demo uses a built-in LIBSVM parser, so there is no data preprocessing step.

[hcho3]

More details: in the test set machine.txt.test, the features 34 and 35 are all missing. So the largest feature ID in machine.txt.test is 33, and XGBoost infers the dimension of the test set to be 34x34.

My hope is that XGBoost should handle this scenario gracefully, i.e. test matrices whose last columns are empty. Rather than throwing error, XGBoost should treat all feature values as missing for features 34 and 35. I believe that's how XGBoost used to behave, since the regression demo was around since XGBoost version 0.60.

[hcho3]

Idea: Use training argument in XGBoosterPredict() so that we can accept LIBSVM test data with fewer number of columns than training data.