Machine Learning for Python
Train a function on training data (x_i,y_i) by minimising the loss: sum_i loss(fx_i,y_i) + lambda * Regularizer (1)
The two main objects in the module are:
- mlmodel: the model to be fitted
- decomposableloss: the loss that should be minimised
The mlmodel defines the function that maps the input to the output, and defines the derivative of that function with respect to its parameters. Next to this basic functionality, also a name and an initialisation can be defined. The loss can be a combination a data loss and a regulariser.
For a simple example, have a look at testml.py
There are situations where formulation (1) does not apply. For instance when the Area Under the ROC curve is optimised. Then the (empirical) loss cannot be decomposed into a sum over losses per training object. In that case you have to define a nondecomposable loss.