Neural network evaluation is not thread-safe

[gkronber]

It seems the evaluation of neural network models is not thread-safe.

The effects can be observed in the partial dependence plot which uses evaluates a model for different inputs concurrently. The bug leads to unplausible charts.

[gkronber]

The processing methods in alglib are indeed not thread-safe

[gkronber]

r15739: fixed bug in NN and NN ensemble using a simple lock for processing an input.

[pfleck]

reviewed r15739: Change works, partial dependence plot values looks sensible now (no more random jitter).

Currently, the locks are performed directly on the alglib.multilayerperceptron and alglib.mlpensemble objects. I suggest we use explicit, private lock objects instead, especially because we have a public getter for those alglib objects.

On the other hand, if someone wants to use the alglib objects manually he/she could use the same object for locking even if the object is also used concurrently within our GetEstimated(Class)Values methods. However, I would rather vote for removing the public getter, because it is currently unused and is an implementation detail which should not be public.

[gkronber]

r16168: incorporated feedback from review, made NeuralNetworkEnsembleModel and NeuralNetworkModel readonly. This is a breaking change to the API

[gkronber]

Thanks for the suggestions for improvements. Please review r16168.

[pfleck]

reviewed r16168: looks good, thanks.

Although making the model readonly (and not publicly accessible) is a breaking API change, I think we can release the ticket since the chance of anyone depending on this is small.

[gkronber]

r16387: merged r15739 and r16168 from trunk to stable.

Ticket can be closed if build and unit tests succeed.