TransferRNN

Workgroup Project at CapoCaccia 2019

TransferRNN framework implements an RNN running on Dynapse chip using CortexControl Python API

Installation

1. Download the nightly build of CortexControl from http://ai-ctx.gitlab.io/ctxctl/primer.html#downloads

2. Clone TransferRNN inside the main CortexControl folder

clone https://github.com/he-xu/TransferRNN.git

3. Prepare the x_projection_train.pkl and state_train.pkl files (by default the script will look for them in TranserRNN/data/)

Running

The framework can run in two modes:

• ctxctl : inside CortexControl (faster but trickier to setup)
• rpyc : in a remote python console using RPyC connection (more stable but slower because of data transfering between python threads)

ctxctl mode

1. Start Cortexcontrol: ./cortexcontrol
2. In the CortexControl console run import TransferRNN.run_rnn_in_ctxctl

Note: Most likely, you will get an error while importing numpy. To fix that, you will have to make sure you have numpy version 1.15.4 (with pip install numpy=1.15.4) in your installation of python and then append the path to site-packages of this python installation inside the TransferRNN/run_rnn_in_ctxctl.py script on line 12.

rpyc mode

0. Install RPyC with pip install rpyc

1. Start CortexControl: ./cortexcontrol

2. In the CortexControl console run PyCtxUtils; PyCtxUtils.start_rpyc_server()

3. Run run_rnn_with_rpyc.py in a python console of your choice (tested in Spyder)

Note: Expect the spike counting phase to be slow. Optimization of this phase is the next framework development step.