Currently I work at optimizing my neural network that I trained with keras. For this, played with several activation functions within the layers and I save the models automatically with model.save() and do an evaluation.
Unfortunately, I forgot the exact setup of my best performing network, but how can I recreate all network information from keras? Of course, I use model.load() and there is helpful model.summary(). But how would I get the activation functions, that were used?
model.summary() just provides information about the network architecture itself:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
... ... ...
_________________________________________________________________
dense (Dense) (None, 64) 288064
_________________________________________________________________
dense_1 (Dense) (None, 128) 8320
=================================================================
Total params: 586,408
Trainable params: 586,408
Non-trainable params: 0
_________________________________________________________________
you can get read layers with model.layers and then use that for finding the weights or activation functions. For example
act_fncs = [l.activation for l in model.layers]