Just a collection of useful abstraction on top of the TensorFlow Python API that make it easier to create complex models in fewer lines of code.

There are two guiding principles:

1. Everything is a subgraph, in the spirit of TensorFlow, exposing named inputs and outputs.
2. All configuration parameters are accessible/modifiable from the top level module.

• I should be able to create an LSTM like this:

X, y = SequenceData(filename="shakespeare.txt")
model = LSTM(input=X, output_dim=128)
# Note at this point X and model.input refer to the same tensor.

loss = SomeLossType(predictions=model.output, labels=y)
trainer = Trainer(model=model, loss=loss, method='SGD', max_iterations=1e6)
trainer.run()

• Starting from pre-trained weights should be a breeze.

• Become well acquainted with the TensorFlow python API. A library like this surely already exists. That's not the point.
• Facilitate exploring methods presented in publications.