numpynet

Convolutional Neural Network written from scratch using numpy with API similar to tensorflow. Library was compared with tensorflow versions of network (demo directory) and achieved very close results.

Implemented Elements

Layers

• InputLayer
• DenseLayer
• BiasLayer
• ActivationLayer (relu, leaky reLu, sigmoid, tanh, sin)
• DropoutLayer
• FlattenLayer
• Conv2DLayer (with bias & stride)
• Pool2DLayer (max, min)
• Padding2DLayer
• Crop2DLayer
• SoftmaxLayer

Losses

• MSE
• CCE

Initializers

• ConstantInitializer
• RandomNormalInitializer
• RandomUniformInitializer
• GlorotUniformInitialization

Metrics

• CategoricalAccuracy

Callbacks

• ModelCheckpoint
• EarlyStopping

Usage Example

Definition

layers = [
    numpynet.layers.InputLayer((28, 28, 1)),
    numpynet.layers.Conv2DLayer(32, kernel_size=3, stride=1),
    numpynet.layers.ActivationLayer('relu'),
    numpynet.layers.FlattenLayer(),
    numpynet.layers.DenseLayer(128),
    numpynet.layers.BiasLayer(),
    numpynet.layers.ActivationLayer('relu'),
    numpynet.layers.DropoutLayer(0.5),
    numpynet.layers.DenseLayer(10),
    numpynet.layers.BiasLayer(),
    numpynet.layers.SoftmaxLayer(),
]

model = numpynet.network.Sequential(layers)

Compilation

model.compile(
    loss='cce',
    metrics=['categorical_accuracy']
)

Fitting

checkpoint_callback = numpynet.callbacks.ModelCheckpoint('checkpoint.dat')

history = model.fit(
    train_x,
    train_y,
    validation_data=(test_x, test_y),
    learning_rate=0.001,
    epochs=10,
    callbacks=[checkpoint_callback],
)

Predicting

predictions = model.predict(test_x)