Word Vector Learner performance

[seoseokho83]

Hi,

I've been trying to run Word Vector Learner on KNIME Analytics Platform 3.3.2. I was wondering how others have found the performance of this node to be, as it is taking forever for mine to work.

I'm trying to run the sample: 06_Calculate_Document_Distance_Using_Word_Vectors workflow from the public server. I've yet to finish running the workflow.

I have a CUDA enabled with GTX 1060 6GB, 24GB RAM, i7-6700HQ. Under preference I have DL4J to use GPU and I can check that the GPU is being used. I also have KNIME use up to 8192m in the ini file.

Any thoughts would be much appreciated.

(FYI: other DL4J workflows work fine)

[seoseokho83]

For some reason, turning off the GPU calculation, allows me to finish the workflow.

[DaveK]

Hi seoseokho83,

unfortunately, all Word Vector nodes will run very slow when GPU is enabled as the Word Vector code in the DL4J library is not optimized for GPU. The reason for this are some implementation specific details of the library which we (KNIME) can't do anything about. Running Word Vector on GPU is officially discouraged by the DL4J developers. So any Word Vector related nodes should be run using CPU only (I think we should a note about that to the node descriptions).

Sorry for the inconvenience.

David

[seoseokho83]

Hi David,

Thanks for the comments. I missed this on the site (shame on me):

Deeplearning4J IntegrationTextprocessing and GPU

Due to the implementation of the textprocessing functionality in the Deeplearning4J library, it is recommended to execute all Deeplearning4J Integration -Textprocessing nodes using the CPU. Performance on GPU may be significantly lower.

 

However, after updating to 3.4 (and the DL4J updated as well), I actually now see that the learner node is actually quite fast even with GPU turned on.