convolutional_network

A simple convolutional network library for image classification. Following layers are made available:

• net.layers.Input
• net.layers.Convolution2D
• net.layers.Flatten
• net.layers.Softmax

Simple example:

import net.layers
import net.model

layers = [
        net.layers.Input(sample_shape=(28, 28, 1)),
        net.layers.Convolution2D(filters=20, rows=14, columns=14),
        net.layers.Convolution2D(filters=10, rows=15, columns=15),
        net.layers.Flatten(),
        net.layers.Softmax()
    ]

model = net.models.Model(layers)

for x_batch, y_batch in x_train, y_train:
    model.train(x_batch, y_batch, learning_rate=0.01)

print("Accuracy: {}".format(model.get_accuracy(x_test, y_test)))

A more comprehensive example can be found in ./examples/mnist.py.

Only 2D convolutions are available and each convolution applies ReLU activation to its output. Padding is not supported, all convolutions are performed in valid mode only. Output of last convolutional layer has to have spatial size 1x1 and channels count equal to number of classes you want to predict. Final two layers must be net.layers.Flatten followed by net.layers.Softma. Simple stochastic gradient descent is used to train networks parameters.

Dependencies

The only dependencies library itself has are python 3 and numpy.
In addition ./examples/mnist.py script uses sklearn to load MNIST data and tqdm for progress bar.

Performance

While this work is no match for commercial grade deep learning libraries, a reasonable amount of effort has been put into optimizing convolutions. All critical computations are performed with matrix dot products.