Core dump due to "out of host memory"

[kevin1kevin1k]

Hi,

I just ran ./thundersvm-train /tmp2/b03902086/data/news20.binary on a workstation machine, and resulted in the following:

2017-12-31 02:14:28,273 INFO [default] training start
.
2017-12-31 02:14:51,947 INFO [default] rho = -0
2017-12-31 02:14:51,947 INFO [default] #sv = 1024
2017-12-31 02:14:52,008 INFO [default] #total unique sv = 1024
2017-12-31 02:15:06,078 FATAL [default] out of host memory
2017-12-31 02:15:06,079 WARNING [default] Aborting application. Reason: Fatal log at [/home/student/03/b03902086/thundersvm/src/thundersvm/thundersvm-train.cpp:90]
Aborted (core dumped)

The machine has 70G memory, and the dataset news20 is just 134M large.
Is there anything I miss?
Thanks in advance!

[shijiashuai]

try ./thundersvm-train -c 4 -g 0.5 /tmp2/b03902086/data/news20.binary

It's a bug here anyway. The default gamma was not been set successfully (so it was 0), and then error occurred.

"out of host memory" comes from the default batch size (10000) of prediction after training. Since news20 has high dimension (1,355,191), 10000 * 1355191 is too large.

• fix default gamma bug
• auto set prediction batch size

[kevin1kevin1k]

Thanks for your quick reply and fixes!