FractalNet

A fractal-based neural network architecture: arXiv paper

Drop-path

I provide a reference implementation for the elementwise layer with global and local drop-path.

Ideal

I use a new layer GlobalDropTrigger as a messager. It sends information to all the FractalJoin in the same fractal block, and tells them drop path as global or local.

Caffe

See the caffe directory for code and more information.

Fractal pattern generation

Wiring up a fractal network manually would take hours, so we provide simple Python scripts that will do it for you. See the generation directory for code and more information.

Data augmentation

We use a Python layer in Caffe to implement data augmentation. It is not yet available here.

Cite

If this is useful to you, please consider citing us::

@article{larsson2016fractalnet,
  title={FractalNet: Ultra-Deep Neural Networks without Residuals},
  author={Larsson, Gustav and Maire, Michael and Shakhnarovich, Gregory},
  journal={arXiv preprint arXiv:1605.07648},
  year={2016}
}