A fractal-based neural network architecture:

• Project page
• arXiv paper

We provide a reference implementation for the elementwise-mean layer with local drop-path. There is still no public release of local+global, but we suggest implementing this through tying weights.

See the caffe directory for code and more information.

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.

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

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}
}