Switching devices to CPU for MaxPooling2DLayer

[m-usama-z]

Hi. For performance reasons, it seems useful to switch devices to CPU for MaxPooling2DLayer, and then switch back to GPU if the original device were GPU. Although, I only observed performance on a few examples so I cannot yet speak for the general case.

[mldiego]

Thanks for catching this! However, it may depend on the hardware (CPU, GPU) and OS used, so we'll keep this in mind, but an analysis over different OS and hardware may be needed. For now, we'll keep it customizable for users to choose the computation device.

[m-usama-z]

Thank you for that information. Instead of the option to use GPU throughout the reachability analysis, it may be more useful to make it an option to use GPU only for convolutional layer (or only for any such layers where it's known to actually improve performance).