DL4J Image Recognition Celebrity Example doesn't work

[sv]

I am attempting to run the celebrity image recognition example and have not made any changes to the workflow. However every time I run it, it produces no errors but the output classes everything as a single class type, in this case it detects everything as ‘angelina jolie’. Then if I append the class prediction probabilities, everything is shown as ‘NaN’. I haven’t changed anything and an using the latest of all nodes and KNIME 3.6.0.

I tried to use a different dataset and replicate it based upon the example but it does the exact same thing.

Is this a known issue?

[stelfrich]

Hi @sv,

@sv

However every time I run it, it produces no errors but the output classes everything as a single class type, in this case it detects everything as ‘angelina jolie’. Then if I append the class prediction probabilities, everything is shown as NaN.

I could reproduce this behavior only once by stopping the learning of the DL4J Feedforward Learner (Classification) before it could even finish just one epoch. And even that doesn’t trigger the observed behavior anymore now.

Have you tried to reset the DL4J Feedforward Predictor (Classification) and re-run it? Does the issue persist?

Best,
Stefan

[sv]

Hi Stefan,

Thanks for your feedback.
When I encountered the issue, I never stopped the learner.
I have discovered a fix for it after a few days. Changing the ‘Dense Layer’ nodes’ Activation Function from RELU to RandomisedLeakyReLU, allowed the process to run completely with filled out prediction probabilities. It could be that something is dropping off in the early stages and causing it to fail.