Repo for focusing neuron (adaptive locally connected neuron)
Paper: https://arxiv.org/abs/1809.09533 Notes:
- Paper is an older version with slightly different focus function normalization
- Current code can provide even better results.
Depends on other libraries: numpy, scikit, theano, lasagne
UPDATE Nov 2019 : Keras version is transferred to another REPO.
Some experiment jupyter notebooks are provided in experiment-notebooks folder.
To run in Google colab you must upload Kfocusing.py and keras_utils.py for Keras based.
Quick example runs on synthetically generated classication datasets: *python Test-Synthetic-Inputs.py
*python mnist.py focused_mlp:2,800,0.25,0.25 10 1 mnist mnist10 0.0
Test set accuracy is ~99.10-99.20
*python mnist.py mlp:2,800,0.25,0.25 10 1 mnist mnist10 0.0
Test set accuracy is ~98.9-99.05
Requires mnist.npz or downloads it from http://yann.lecun.com/exdb/mnist/ Other datasets such as cifar_10 and fashion can be downloaded with keras.datasets Note: mnist_cluttered data is difficult to find in internet again. Email me if you cant find it. I will upload it
Repeated trial experiments are implemented .sh files. Contains my local directory references.
UPDATE AUG 2019:
I have added keras implementations and some new ipynb for experiments:
Kfocusing: the focusing neuron layer class file, include a unit test (Requires included keras_utils.py)
KfocusingTransder: test focusing neuron in transfer learning with keras.applications and pretrained models (VGG1-16)
Boston experiment notebook
Reuters experiment notebook (however, theano and python version worked better)
Random_Syntethic_Tests-master-forPaper-ready.ipynb repeats the synthetic experiments
Focusing_Network_Test_Single_Run-Mnist-For-Paper-ready.ipynb repeats a single run MNIST experiment
USE dense-nn-weights-mnist-eng-ready.ipynb to experiment on Dense network Weights with noise PADDED MNIST
NOTE Keras versions can be run in GOOGLE colab