Strip off sunglasses! : Discrete Hopfield Network

Summary

This Python code is just a simple implementaion of discrete Hopfield Network (http://en.wikipedia.org/wiki/Hopfield_network). Discrete Hopfield Network can learn/memorize patterns and remember/recover the patterns when the network feeds those with noises.

Example (What the code do)

For example, you input a neat picture like this and get the network to memorize the pattern (My code automatically transform RGB Jpeg into black-white picture).

After the network memorized it, you put the picture with noise(sunglasses) like this into the network.

The network can strip off the sunglasses because the network remembers the former picture.

Instructions

##Input files## JPEG files like those in "train_pics". Network learns those pics as correct pics. If you want to add new pics, please put them in "train_pics" folder.

##Test files## The pictures with sunglasses should be in "test_pics" folder.

##How to run the code## Prior to running my code, please install the following libraries.

• numpy
• random
• Image
• os
• re

Also, Here is the way to run my code.

• Theta is the threshold of the neuron activation.
• Time is a parameter telling the steps of remembering the learned pictures. As the number of the steps increases, the remembered picture is more accurate.
• size is the picture size in pixel. If you put a pic with different sizes, the code resize it.
• threshold is the cutoff threshold to binarize 1 byte (0 to 255) brightness.
• current_path should be current working folder path (usual way is os.getcwd())

After you download all the files in this repository, please run "hopfield.py". Here is the main code.

#First, you can create a list of input file path
current_path = os.getcwd()
train_paths = []
path = current_path+"/train_pics/"
for i in os.listdir(path):
    if re.match(r'[0-9a-zA-Z-]*.jp[e]*g',i):
        train_paths.append(path+i)

#Second, you can create a list of sungallses file path
test_paths = []
path = current_path+"/test_pics/"
for i in os.listdir(path):
    if re.match(r'[0-9a-zA-Z-_]*.jp[e]*g',i):
        test_paths.append(path+i)

#Hopfield network starts!
hopfield(train_files=train_paths, test_files=test_paths, theta=0.5,time=20000,size=(100,100),threshold=60, current_path = current_path)

Reference

• http://en.wikipedia.org/wiki/Hopfield_network
• http://rishida.hatenablog.com/entry/2014/03/03/174331 (in Japanese)

---