Caffe is a deep-learning framework created by Yangqing Jia during his PhD at the University of California, Berkeley, and maintained by the Berkeley Vision and Learning Center (BVLC, later BAIR). Begun around 2013 and described in the 2014 paper “Caffe: Convolutional Architecture for Fast Feature Embedding” (Jia et al.), it was one of the first widely used frameworks built specifically for training and deploying deep convolutional neural networks, and it became a default tool for computer-vision research in the mid-2010s.
Engineering-wise, Caffe was a C++ library with CUDA GPU kernels, wrapped with Python and MATLAB interfaces. That design put speed first: the official site advertises processing of tens of millions of images per day on a single GPU, on the order of a few milliseconds per image. By writing the heavy numerical work in optimized C++ and CUDA while exposing a scriptable surface for experiments, Caffe gave researchers performance without forcing them to work entirely in a low-level language.
Its most distinctive design choice was declarative model definition. Rather than writing code that built a network programmatically, a Caffe user described the network as a configuration: layers, their types, and their connections were specified in a plain-text protocol-buffer file, and a separate solver file described the training schedule. This decoupling of model representation from implementation, emphasized in the paper, meant the same model definition could be moved between CPU and GPU, shared, and version-controlled as data rather than code.
The Model Zoo turned that declarative format into a community asset. Because a trained model was captured as a standard definition plus a weights file, BVLC could publish reference models and others could share their own. Researchers could download a network such as a pretrained ImageNet classifier and either run it directly or fine-tune it, which made Caffe an early hub for transfer learning in vision and helped spread the practice of building on published pretrained networks.
Caffe was released under a BSD license and attracted a large contributor community. It was eventually superseded by the more flexible, Python-centric frameworks that followed, and its lineage continued in a successor effort (Caffe2) aimed at production and mobile that later merged into the PyTorch project. As one of the first fast, GPU-accelerated, openly distributed deep-learning frameworks, Caffe played a formative role in the computer-vision boom and in establishing pretrained model sharing as a standard practice.