NumpyNets

A modular, keras like, implementation for building neural networks using Numpy.

Figure 1: Visualization of the feed forward network's output. The circle indicates the non-linear decision boundary of the net.

Regression:

regression_net=feed_forward(input_dim=5, task='regression', non_linearity='tanh')
regression_net.add(10) #add a layer with 10 neurons, plus a bias
regression_net.add(8)
regression_net.add(1) #output
regression_net.predict(np.random.rand(1,5)) #make a prediction
regression_net.train_network(x,t,iterations=100) #train

Classification:

class_net=feed_forward(input_dim=10, task='classification', num_classes=3, non_linearity='relu')
class_net.add(10)
class_net.add(3) #number of classes
class_net.predict(np.random.rand(1,10)) #predict the probabilities.