Monstruous deep learning framework built on top of Theano and other frameworks such as blocks, fuel and lasagne.
The main motivation was that doing RNN and extending them in lasagne is a bit for now cumbersome. Adding batch norm and converting their LSTM to a ConvLSTM meant basically to copy paste the whole class and hard code the changes. It also seemed like a more rewarding challenge to code a framework than to try to crack open the blocks bricks obscure dark magic.
- From fuel it uses pretty much all the data handling part
- From blocks it uses the MainLoop class (extensions are heavily used!) and the Algorithm classes
- From adlf is the core framework for building neural nets
- From lasagne it is pretty much a toolbox of added stuff (like batch norm and weight norm) when not using adlf but going more for straight lasagne.