nnetsys is a library enabling very easy creation, teaching, and testing of neural networks.
For example, if you wanted to train a simple DNN, you would just type:
from nnetsys.ff import Network, PerceptronLayer, MinibatchSGDTeacher
net = Network(
PerceptronLayer(500, 300, activation='relu', dropout=0.5),
PerceptronLayer(300, 200, activation='relu', dropout=0.5),
PerceptronLayer(200, 50, activation='relu', dropout=0.5),
PerceptronLayer(50, 10, activation='softmax')
)
teacher = MinibatchSGDTeacher(net, training_data, batch_size=10, learning_rate=0.1)
for i in xrange(0, 100):
teacher.train_epoch()It can't get much simpler, can it?
nnetsys supports:
- multilayer perceptrons (sigmoid, tanh)
- deep neural networks (relu + dropout)
- convolutional neural networks
Under the hood nnetsys uses Theano, so it runs pretty damn fast.