![[Torch]](images/title.jpeg)
![[XXXXX]](images/torche.jpeg)
You
should have a look on this presentation
written by Samy Bengio
which introduces main concepts in Torch.
Here is an abstract:
There are several concepts to know in Torch, if you plan
to use it:
-
DataSet
-
Machine
-
Trainer
-
Measurer
A DataSet is a black-box which could provide all kind of
data. For example you have:
-
StdDataSet, which manages standard datasets: examples with the same input
dimension and the same output dimension.
-
SeqDataSet which provides a easy interface to sequence datasets used in
the Distribution package...
A Machine is a learning-machine. It could be:
-
MLP, a multi-layered perceptron.
-
HMM, a hidden markov model
-
SVM, a support vector machine
A Trainer is a class which takes a Machine and a DataSet,
and which is able to train or test the machine. It could be:
-
GMTrainer, a gradient machine trainer which can train all kind of gradient
machines. (For example MLP)
-
EMTrainer, which trains several machines using the EM algorithm. (For example
HMM)
-
QCTrainer, which trains all quadratic, constrained machine. (For example
SVM)
A Measurer is a special class which is called by a Trainer,
during the train phase or the test phase. It's always associated to a DataSet.
It could measure everything on your machine, on the DataSet that
you provide. Here are some examples:
-
MseMeasurer, which measures the mean squared error of you machine, on the
DataSet.
-
ClassMeasurer, which measures the classification error when you have a
classification problem.
Ok, almost all classes in Torch are one of these classes...
Easy, no ? If you want to have an idea of all classes which are available
to the public, just have a look on the reference
manual...