batch-size

[gcastro-98]

Implement a batch_size argument for the training method.

Currently, the gradient descent is purely stochastical (batch_size=1).