neural-networks

A suite of neural-network architectures, training algorithms and useful utility functions written in python.

A short example usecase (data not actually provided):

from network import NeuralNet
from backprop import train
from layer import LogisticLayer, LinearLayer, SoftMaxLayer
from metrics import error

#to load matlab format data
from scipy.io import loadmat

mnist = loadmat('MNIST60k.mat')

training_data = mnist['train']
validation_data = mnist['valid']

model = NeuralNet([LinearLayer(784), LogisticLayer(1000), SoftMaxLayer(10)])

epochs = 1000

#Train the model using backprop
for epoch in range(epochs):
    
    #do a single iteration of backprop
    train(model, training_data['input'], training_data['target'])
    
    #evaluate on validation set
    valid_error = error(model, validation_data['input'], validation_data['target'])
    print 'EPOCH %d VALID CROSS ENTROPY %.5e' % valid_error