DL4J Feedforward Predictor fill out the /tmp directory

Hello,

I am using in Knime the "DL4J Feedforward Predictor (Classification)" in parallel (with parallel chunk) for a lot of MLP models on a lot of images.

I am running Knime 3.4.1 on Centos 7.

Each DL4J Predictor node creates in the /tmp directory a restoreXXXXXmultiLayer file where XXXXX is some integer.

There are 2 problems :

  • This restoreXXXXXmultiLayer file should be created in the temp directory defined in Knime preferences, currently it's not ; it goes to /tmp whatever is in the knime preferences.
  • This file is not deleted at the end of the node execution

Hence /tmp explodes the root partition.

I don't have access to the OS configuration and the other users on the system are beginning to be a bit agressive :-).

I have done some post node execution cleaning in /tmp but it would be nice to have this corrected (in 3.4.2 ?).

Thanks

Philippe

 

Hi Philippe,

thank you for reporting that issue. I'll have a look and try to fix the problem.

Cheers

David

Hi Philippe,

sorry for the long wait. Unfortunately, these files are created by the DL4J library. Therefore, we can't control the loaction of the tmp dir were they are saved to. Currently, they should be deleted after KNIME closes. Could you maybe confrm that these indeed get deleted after you close the application?

However, this won't help if you have a very long running loop. I'll open an issue on the DL4J Githup page and ask them to change the behaviour.

Cheers

David