A small and pure Numpy Convolutional Neural Network library I wrote in order to understand backprop through conv layers.
Here is an example on how to build a not-so-deep convnet that uses strided convolution:
layers = [ Conv((4, 4, 1, 20), strides=2, activation=lkrelu, filter_init=lambda shp: np.random.normal(size=shp)), Conv((5, 5, 20, 40), strides=2, activation=lkrelu, filter_init=lambda shp: np.random.normal(size=shp)), Flatten((5, 5, 40)), FullyConnected((5*5*40, 100), activation=sigmoid, weight_init=lambda shp: np.random.normal(size=shp)), FullyConnected((100, 10), activation=linear, weight_init=lambda shp: np.random.normal(size=shp)) ] net = Network(layers, lr=0.001, loss=cross_entropy)You can see that it somehow takes inspiration from Keras.
- Python 2.7
- Numpy