Problems with multithreading and generation step

[BenoitMiquey]

Hi,

I realy appreciate your works on PyGAD!

I'm using it to make some chaotic learning with thousands of model, and a greedy fitness function. the parallelization is realy efficient in my case.

I have found some problems with multithreading using keras models.

To reproduce the problem, i use this regression sample : https://pygad.readthedocs.io/en/latest/README_pygad_kerasga_ReadTheDocs.html#example-1-regression-example

I only reduce the num_generations to 100.

Steps to reproduce :

I run a few times the sample,

• then, i enable the parallel processing on 8 threads :

• then, run again a few times :

• sometimes, i see in logs a fitness lower than the n-1 generation, example :

• I printed all solutions used in each epoch, and i saw thats solutions are most of time the same, so the parallel_processing seems to break the generation of the next population in the most of cases.

Thanks!

EDIT :
In addition i tried to reproduce the same problem with this classification problem sample ,
Adding the multiprocessing support cause the same problem.

[ahmedfgad]

Thanks to https://stackoverflow.com/a/75606666/5426539.

Keras is not thread-safe. The issue can be simply solved by cloning the model before calling the predict() method.

model = ...
_model = tensorflow.keras.models.clone_model(model)
predictions = _model.predict(...)

PyGAD 3.0.0 solved this issue.