Can GPU be used during trainning?

[lk1983823]

Can GPU be used during trainning such as XGBOD ? Thank you !

[yzhao062]

Hi Frank,

Unfortunately, most pyod models are not compatible with GPU computation. XGBOD itself actually would not benefit from GPU too much as the most expensive part is generating additional features, i.e. running a few unsupervised outlier detection algorithms. We did consider to enable GPU: https://pyod.readthedocs.io/en/latest/todo.html in the future, but for now it is still on the todo list and not decided when/how to do it efficiently.

Best, Yue

[lk1983823]

Thanks a lot. However, I wonder whether this tool can be used under big data enviroment in the future?

[TimotheeGr]

I am using the MO-GAAL and I was hoping to benefit from GPU but it is only 25% faster for the training. Is it expected? I am wondering if I am doing something wrong somewhere ^^

[AmalMtibaa]

I am using the MO-GAAL and I was hoping to benefit from GPU but it is only 25% faster for the training. Is it expected? I am wondering if I am doing something wrong somewhere ^^

Hi TimotheeGr, can you tell me how you enabled training on GPU for MO-GAAL ?

[yzhao062]

if you are using the old version tensorflow (<=1.15), then make sure you installed tensorflow-gpu. for newer version tensorflow, that is automatic.

Given GPU is being used (search google "tensorflow not using gpu" if not), mo_gaal's usage is automatic. We get the consent from the paper author so that is based on their code.

[TimotheeGr]

After my question above I checked that the GPU was used with nvidia-smi and it is only used at 7-10% level which probably explains why it is not a big improvement. (But I am quite a beginner on this)