❗❗ NOTE: Here is the link to our new project repository, which is much better documented and contains more comprehensive experiments. The best part is, it's written in JAX so it's also much easier to follow!
This repository contains a set of experiments written in Tensorflow (version 1.x) to explore the node perturbation algorithm for training deep fully connected neural networks.
Run master.py to train a multi-layer perceptron (with sgd or np).
python master.py -lr 0.1 -update_rule np -n_hl 3 -hl_size 300 -n_epochs 5All are optional arguments:
lr: learning rate
update_rule: either 'np' or 'sgd'
network will have (n_hl) depth and constant width of size (hl_size)