Decaf is a framework that implements convolutional neural networks, with the goal of being efficient and flexible. It allows one to easily construct a network in the form of an arbitrary Directed Acyclic Graph (DAG) and to perform end-to-end training.

For more usage check out the wiki. A great place to start is running ImageNet classification on an image.

For the pre-trained imagenet DeCAF feature and its analysis, please see our technical report on arXiv. Please consider citing our paper if you use Decaf in your research:

@article{donahue2013decaf,
  title={DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition},
  author={Donahue, Jeff and Jia, Yangqing and Vinyals, Oriol and Hoffman, Judy and Zhang, Ning and Tzeng, Eric and Darrell, Trevor},
  journal={arXiv preprint arXiv:1310.1531},
  year={2013}
}