Caffe

What Caffe is

Caffe is a library for training and deploying convolutional neural networks (CNNs) and other deep models. It is written in C++ and released under a BSD license, with bindings for Python and MATLAB: a user writes scripts in those languages, while the computation itself runs in compiled C++ code.^[1] A convolutional network is a stack of layers whose filters are learned from data rather than written by hand, and the library's job is to take a description of such a stack, run the forward and backward passes over training data, and then run the trained network on new inputs. The BSD license is a permissive open-source license that allows use, modification and redistribution, which matters to teams that want to build the library into their own systems.

Where it came from

The paper describing Caffe, "Caffe: Convolutional Architecture for Fast Feature Embedding", was first posted to arXiv on 20 June 2014.^[1] Its authors were Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama and Trevor Darrell.^[1] The code was maintained and developed by the Berkeley Vision and Learning Center, with help from contributors on GitHub: the lab kept the core of the code, and outside users sent fixes and additions.^[1]

What it is good at

Speed is the paper's emphasis. Caffe uses CUDA for computation on GPUs, and the paper reports that a single K40 or Titan GPU can process over 40 million images a day, about 2.5 milliseconds per image.^[1] That number describes throughput in a vision pipeline, where images pass through a network one after another.

A second design choice is that the framework separates model representation from implementation, so the same model can move from a prototyping machine to a cloud environment without being rewritten.^[1] Caffe also ships with a collection of reference models for state-of-the-art deep learning algorithms, which means a user can start from an existing network description rather than an empty file.^[1] The authors say the library powers research projects, large-scale industrial applications and startup prototypes in vision, speech and multimedia.^[1]

Its place in the field

Taken together — a configuration-based description of a network, a set of reference models, and GPU throughput measured in milliseconds per image — these choices lower the cost of reproducing and comparing results. When the network is a file rather than code, the same file can be handed to someone else and run on their machine, and a published model gives people a common starting point instead of each group rebuilding the same thing.

The phrase in the paper's title, "fast feature embedding", points at the same priority: turning images into feature representations quickly, so that recognition and other tasks can be built on top of those representations. Caffe's stated scope is convolutional networks and other deep models, and the applications its authors name are in vision, speech and multimedia — research, industrial deployment and startup prototypes rather than a single use case.^[1] The library's identity therefore comes from a combination of things that are easy to describe separately but useful together: a permissive license, language bindings that let researchers work in Python or MATLAB, a model format that survives the move from a laptop to a cluster, reference models to start from, and throughput reported in images per day on a single GPU.^[1]