v0.6.1

Version 0.6.1 of theanets is now live!

pip install -U theanets

http://pypi.python.org/pypi/theanets
http://theanets.readthedocs.org
http://github.com/lmjohns3/theanets

The biggest change in this release series is a Network/Layer refactor that preserves the existing API but permits much more flexible network layouts if desired. Layers can now output multiple values; by default most layer types generate an "out" (the traditional layer output) as well as a "pre" (the layer's pre-activation value). Other notable changes include:

• The semantics of the "rect:min" and "rect:max" activations has been reversed -- rect:min now gives g(z) = min(1, z) and rect:max now gives g(z) = max(0, z). The "relu" activation still means g(z) = max(0, z).
• Theanets now uses Travis CI and Coveralls.io to build and compute test coverage automatically -- see https://travis-ci.org/lmjohns3/theanets and https://coveralls.io/r/lmjohns3/theanets. Test coverage increased from 76 to 91%.
• The documentation has been expanded and hopefully made more clear. There's always more room for improvement here!
• Activation functions are now first-class objects. New activation functions include Prelu, LGrelu, and Maxout.
• Loading and saving uses the standard pickle module.
• Almost all of the trainers have moved to a new package, see http://downhill.readthedocs.org.

As a reminder, the 0.7.x release series will incorporate several big changes, but most important is that recurrent models will reorder the axes for input/output data; see goo.gl/kXB4Db for details.

As always, I hope the library will be really useful! Please file bugs, post on the mailing list, etc. as you run into questions or issues.

Version 0.5.0

Version 0.5.0 of theanets is now live!

pip install -U theanets

http://pypi.python.org/pypi/theanets
http://theanets.readthedocs.org
http://github.com/lmjohns3/theanets

Some great new features have been incorporated into this release, but
the biggest one is that Layers have been refactored into first-class
citizens. It's much easier to specify model layers, and many different
types of recurrent layers are now available.
http://theanets.readthedocs.org/en/stable/creating.html#specifying-layers

I've also tried to improve the speed of the models and trainers, and I
have some ideas that I'll be incorporating into future releases in
this area.

I've tried to get the documentation into better shape. It still needs
some work, but it's a bit better than it has been.

This release also includes code from 4 first-time contributors!

Please note that this version makes several backwards-incompatible
changes that I think will be a net improvement, at the cost of
potentially breaking some of your existing training scripts. Most
notably:

• The code relies on a new release of the climate package. Be sure to
  install using "pip install -U theanets" to get the most recent
  dependencies.
• The Experiment.itertrain() method now generates two dictionaries
  of monitor values: one for training, and one for validation.
  http://theanets.readthedocs.org/en/stable/training.html#training-as-iteration
• A Network now has a find() method for retrieving shared variables
  (e.g., weight matrices); the get_weights() method has been removed.
• Trainer.train() has been renamed Trainer.itertrain(), and the
  SGD-based trainers have been refactored a bit, so now there is no
  longer an SGD.train_minibatch() method.

I hope the library will be really useful! Please file bugs, post on
the mailing list, etc. as you run into questions or issues.