WayneZww / FractalNet

Use the FractalNet of caffe version from gustavla. Add the Windows PowerShell scripts and python scripts to generate the prototext file for caffe with fixing the encoding problem.

FractalNet

Use the FractalNet of caffe version from gustavla <https://github.com/gustavla/fractalnet>__. Add the Windows PowerShell scripts and python scripts to generate the prototext file for caffe with fixing the encoding problem.

A fractal-based neural network architecture:

• Project page <http://people.cs.uchicago.edu/~larsson/fractalnet/>__
• arXiv paper <https://arxiv.org/abs/1605.07648>__

Drop-path

The auther provide a reference implementation for the elementwise-mean layer with local drop-path. This project is trying to realize local+global, by tying weights in blocks and globals.crox.

Caffe

See the caffe directory for code and more information.

Fractal pattern generation

Use the provided simple Python scripts, and fixed some problems. See the generation directory for code and more information.

Data augmentation

Try to use the lmdb file as input layer in Caffe to implement data augmentation. The way to build a cifar10 dataset for caffe can been see

reference

Please citing ::

@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}
}