nnutil

Tool for creating numpy neural nets

A little module for quickly and simply creating numpy neural nets. Found useful for eliminating dependencies for quick poc/demo purposes where "this thing needs to run on this laptop in three hours and TensorFlow just broke". If you've got a static, stable environment you should probably use something else.

Quick training example:

df = pd.read_csv('example.csv')
costs = []
it = minibatch_gen_from_pddf(df, "Y (Target Label)", 1024)
for X, Y in it:
    if not costs:
        layers = [[16, "relu"], [8,"relu"], [1, "sigmoid"]]
        nn = NN(layers=layers, data=X.T, labels=Y, learning_rate=0.002)
    else:
        nn.data = X.T
        nn.labels = Y
    costs.append(nn.train(10000)[0])

After training classification can be done simply by calling model_forward of the instance, passing new data.