Getting started

Linear perceptron

You can replicate a linear regression (OLS) with a MLP with no hidden layer and a linear output activation.

from statinf.data import generate_dataset
from statinf.ml import MLP, Layer

# Generate the synthetic dataset
data = generate_dataset(coeffs=[1.2556, -6.465, 1.665414, 1.5444], n=1000, std_dev=1.6)
new_df = generate_dataset(coeffs=[1.2556, -6.465, 1.665414, 1.5444], n=500, std_dev=1.6)

Y = ['Y']
X = [c for c in data.columns if c not in Y]

# Initialize the network and its architecture
nn = MLP(loss='mse')
nn.add(Layer(4, 1, activation='linear'))

# Train the neural network
nn.train(data=data, X=X, Y=Y, epochs=1, learning_rate=0.001)

# Predict new output
pred = nn.predict(new_data=new_df)

# Get the network's parameters
print(nn.get_weights())

The output will be:

{'weights 0': array([[ 1.32005564],
   [-6.38121934],
   [ 1.64515704],
   [ 1.48571785]]), 'bias 0': array([0.81190412])}

Non-linear model

You can fit complex data structures by stacking several layers. Keep in mind that the input dimension of on layer corresponds to the output size of the previous (except for the first layer, having the number of features as input dimension).

The below example shows how to fit a complex binary classification problem.

from statinf.data import generate_dataset
from statinf.ml import MLP, Layer

# Generate the synthetic dataset
data = generate_dataset(coeffs=[1.2556, -6.465, 1.665414, 1.5444], n=1000, std_dev=1.6)
new_df = generate_dataset(coeffs=[1.2556, -6.465, 1.665414, 1.5444], n=500, std_dev=1.6)

data['Y'] = data['Y'].map(lambda x: 1. if x > 0. else 0.)

Y = ['Y']
X = [c for c in data.columns if c not in Y]

# Initialize the network and its architecture
nn = MLP(loss='binary_cross_entropy')
nn.add(Layer(4, 8, activation='relu'))
nn.add(Layer(8, 2, activation='relu'))
nn.add(Layer(2, 1, activation='sigmoid'))

# Train the neural network
nn.train(data=data, X=X, Y=Y, epochs=100, learning_rate=0.001, optimizer='adam')

# Predict new output
pred = nn.predict(new_data=new_df)

Run STATINF with GPU

See theano documentation for more details and examples.