Antipasti-TF

Antipasti-TF is a lightweight wrapper around Tensorflow for building convolutional neural networks with complex architechtures. In under active development and not ready for deployment - yet. :-)

Planned API

The API is non-functional and uses NetworkX behind the scenes, which allows for arbitrarily complex architectures.

For the following, let's assume that our model is called network.

• We implement an API resembling python lists/tuples for sequentially stacking layers: network = conv(...) + pool(...) + conv(...) + .... This supports constructs like list comprehensions and reductions. In addition, we define a * operation to stack layers laterally, so you could define an inception module like: previous + conv(...) * conv(...) * conv(...) * pool(...) + next. For predefined sub-networks as modules, a U-Net could look like: network = module_1 + (module_2 + module_3 * id() + module_4) * id() + module_5. Slicing would be defined as expected, e.g. subnetwork = network[3:5] and network surgery would be as simple as network[5:8] = another_network[3:5].

• Simultaneously, we expose as much of the graph as possible, so you could do things like network.add_nodes(['conv1', 'pool1', ..., 'conv10', ...]) followed by network.add_connection(from_layer='conv1', to_layer='conv10', join_by='concatenation'), or simply network.add_layer(conv(...), previous_layer='pool8').

Why not Keras, a mature library with eloquent code?

If you're reading this, we assume you're somewhat familiar with Keras. Without further ado, there are four reasons:

• Keras is functional for the most part. The functional API can be powerful for constructing models, but altering a model after it has been built is usually a tedious process. We're aiming for a non-functional API, one that affords flexibility before and after the model has been built. This enables efficient net surgery (as it's fondly called in the Caffe community) on pretrained models. Think of it as a keras Sequential model, except it doesn't need to be sequential.

• Keras is a multi-framework tool, wrapping both Theano and Tensorflow. Theano and Tensorflow are awesome libraries, but we have found that supporting both simultaneously does neither justice while slowing down the development process. With Antipasti, we're shooting for a library that fully integrates with Tensorflow (including graceful and transparent handling of data- and model-parallel multi-GPU training (synchronous and asynchronous) out of the box, and full support for distributed tensorflow).

• Keras is a general purpose framework, but Antipasti is being built with images and volumes as first class citizens. A fully connected layer is just a convolutional layer with a 1x1 filter.

• Antipasti does not intend to replace Keras. In fact, it should add to it by implementing a multi-GPU/distributed training bench for Keras models via a Keras to Antipasti compatibility layer.

Getting Involved

Help shape this project! Have suggestions on how the API should look, or time to contribute? Get in touch by email (nasim.rahaman at iwr.uni-heidelberg.de) or opening an issue.

Who's involved?

As of the day:

• Nasim Rahaman of Image Analysis and Learning Lab @ Heidelberg Collaboratory for Image Processing.