Example of DeepESN and NG-RC about reservoirpy HOT 3 CLOSED

Hello,
We are indeed a bit late on the documentation of the new beta version, but it is on the way !
About the NG-RC, you can already find a pretty complete reproduction of the original paper in the examples/ folder, in the form of a notebook.
About the DeepESN, you can try them out using ReservoirPy Node API as follow (for a 3 layers model for instance):

from reservoirpy.nodes import Input, Reservoir, Ridge

inputs = Input()

r1 = Reservoir(...)  # the parameters are up to you
r2 = Reservoir(...)
r3 = Reservoir(...)

readout = Ridge(...)

deep_esn  = (inputs >> r1 >> r2 >> r3) & ([r1, r2, r3] >> readout)

This should produce the following architecture:

r3------
|       |
r2---- concat  -----> readout
|       |
r1------
|
inputs

However, be warned that the Node API is still in heavy development. We would of course be happy to have feedback from you regarding its capabilities.

from reservoirpy.

0.3 is great....huge improvement in disk space usage, memory leaks with new features. All the problems I had are gone.

Are feedback loops possible? I didn't see anything about feedback loops in the papers regarding DeepESN but it seems to be in your code.

Do you have any tips on finding the right parameters for DeepESN especially number of layers?

from reservoirpy.

We are really glad you like it !
Feedback loops are possible using the link_feedback(receiver, sender) function from reservoirpy.ops module, or using the << operator:

node_fb = node1 << node2  # node_fb  is a copy of node1 receiving feedback from node2

node1 <<= node2  # same operation but in place in node1

About the number of layers in a DeepESN I admit I never tried to optimize this kind of model. You can probably find more information in the 2018 paper from Gallichio et al. "Design of deep echo state networks".

from reservoirpy.