[question] load keras with custom_objects about snn_toolbox

[rbodo]

Hi Bjoern,

Thanks for your detailed question.

You are right, the support for custom Keras objects is still a bit of a stump. What you propose should enable you to load the model, with a few caveats:

• The line you quote is in a branch that is entered only when the model is saved as json file, with weights separate. This branch is there for backwards-compatibility for old Keras versions (pre v1), and I'm not even sure Keras still provieds this way of saving/loading models any more. So you could use the second branch (loading the h5 file with custom objects), by linking your custom objects in this dict (despite its name, the function handles any custom objects, not just activations).
• If loading your model was successful, you still need to register these custom layers in the config_default file, otherwise the parser will skip them.
• Finally, (assuming you want to use these layers in the converted SNN), you should implement the spiking versions if those layers here.

Good luck!