Closed HenKlei closed 4 years ago
It should also be possible to determine the architecture of the neural network in the reduce
method. Thus, it is possible to try different architectures without having to create different NeuralNetworkReductor
s. However, the reduce
method should compute the reduced basis only once (this also means that the tolerances/reduced basis size have to be set when creatig the NeuralNetworkReductor
) and reuse the same reduced basis when it is called again.
The first part is done in 3bf98d4.
Second part done in 983df82.
Should parameters like
optimizer
,epochs
,batch_size
,learning_rate
andrestarts
be fixed when creating aNeuralNetworkReductor
or should they be added to the signature of thereduce
method? Maybe one wants to callreduce
multiple times with different training parameters?