One-Class SVM showed out of host memory error about thundersvm HOT 4 CLOSED

@mhwong2007 It's difficult for us to reproduce the problem, because we don't have the dataset and the machine with such large memory. A guess is that "-n 0.5" (meaning half of the training instances are support vectors) may be too large.

Would you please attach the full log for us to try to locate the issue? Alternatively, you can use publicly available datasets to produce the issue again, and then we can debug the code.

from thundersvm.

I will try using a smaller -n option

These are the logs:

2018-01-03 16:59:18,495 FATAL [default] out of host memory
2018-01-03 16:59:18,496 WARNING [default] Aborting application. Reason: Fatal log at [/dev/xvda2/thundersvm/src/thundersvm/thundersvm-train.cpp:97]

from thundersvm.

@mhwong2007 The problem comes from dataset loading. I've just pushed a new version. You may try it to see whether your problem has been solved.

from thundersvm.

Problem still exists...
Here is the log

2018-01-15 07:18:40,994 INFO [default] loading dataset from file "/dev/xvda9/scaled.libsvm"
2018-01-15 07:56:21,231 INFO [default] #instances = 3000000, #features = 3000
Segmentation fault (core dumped)

I used this command to build a linear one class svm

./thundersvm-train -s 2 -t 0 -n 0.1 scaled.libsvm largedataset.model

from thundersvm.