This repository contains the experimental framework, developed by David Warde-Farley, Pascal Lamblin, Ian Goodfellow and others during the winter 2011 offering of IFT6266.
- Models should implement the Block interface.
- Methods on these models should expect Theano expressions as arguments and return Theano variables (except in a few select cases where the modules are trained once and never, ever fine-tuned, for example the K-Means module)
- Parameters should be Theano shared variables.
- Other than the shared variables, the methods should have no side effects in the form of storing intermediate state.
- Penalties and whatnot should be generally kept separate from the actual model class, and implemented as subclasses of the "Cost" object. We break this rule currently in the case of ContractingAutoencoder because the model is intimately tied to the penalty, and just have a contracting_penalty() method that returns the relevant bits to be added to the cost function.
- Subscribe to the pylearn-dev Google group for important updates.