MAPTrainer is a Python framework based on the PyTorch library1
dedicated to machine learning research in the context of MultiAgent
Patrolling (MAP). As PyTorch, it uses GPUs and CPUs. MAPTrainer is a
native Python package by design, called maptrainer. Its
functionalities are built as Python classes, which enable the
integration of its code with Python packages and modules. Relying on
PyTorch, it enables GPU-accelerated training, through CUDA, for machine
learning, and especially neural network applications. The objective of
this framework is to train any machine learning model, built with
PyTorch, for any MAP scenario. Parameters defining the MAP scenario are
provided to the programme.
maptrainer.model allows implementing any machine learning model in the
same way as PyTorch. Any new machine learning model which is intended to
be experimented shall be added in this package. Then, the model of this
class will be loaded by maptrainer.model.modelhandler, which will
check, if specified by the user, whether a previous trained version of
this model for the current MAP scenario exists. Otherwise, a new model
is created for the current scenario.
This package adds support for loading MAP data necessary to train a
model. The data of a MAP scenario are loaded by means of a data
loader, which is merely an object of
maptrainer.data.MAPDataLoader.MAPDataLoader type. MAPDataLoader is
an abstract class defining a template for any data loader. A new data
loader shall extend this class and implements its methods. Over the
training stage, the data loader will then return data divided in
validation and training data, and for each type of data, into input and
output data. In execution, the data will then be returned fold after
fold if a
-
BPDataLoader: loads binary sequences, -
IPDataLoader: loads on-vertex idleness sequences.
maptrainer.train is used to automatically train a model. A complete
training is run with the maptrainer.train.run_epochs method, which
runs a training for the loaded model, the criterion to optimise, the
data loader, the learning rate, and the number of epochs, among others.
maptrainer.valid is used to validate the model with respect to the
criterion to optimise, the data loader, the learning rate, and the
number of epochs, among others.