Caffe

Caffe was developed by Yangging Jia while he was pursuing a Ph.D. at the University of California, Berkeley. Caffe has become popular and it has wide support from a community working in academia, start-ups, large enterprises, media, and others.

One of its big supporters is Yahoo. They integrated Caffe with Apache Spark to create CaffeOnSpark, a distributed deep learning framework.
This combination of products provides them with tremendous processing speed, making it ideal for image processing. In 2017, Facebook launched Caffe2 with features like Recurrent Neural Networks. Eventually, Caffe2 was merged into PyTorch in 2018.

Project Background

Framework: Caffe
Author: Yangqing Jia
Released: April 2017
Type: Open-source library for deep learning
License: BDS
Supports: CPUs, GPUs, Nvidia CUDA, and cuDNN Library (CUDA Deep Neural Network)
Language: C++ and Python
GitHub: BVLC/caffe
Runs on: Mac OS, Windows, Linux, Ubuntu, Red Hat, and OS X

Applications

Classification
Regression
Clustering
Dimensionality reduction
Structured prediction
Anomaly detection
Reinforcement learning
Artificial neural network
Image classification
Image segmentation
CNN, RCNN, and LSTM
Fully connected neural network designs
Large-scale industrial applications in vision, speech, and multimedia

Summary

Supports image classification and image segmentation
Supports CPU (Intel MKL) and GPU (cuDNN)
Supports Region-based convolutional networks (R-CNN) and Faster R-CNN
Popular in academic research projects, prototypes, and large-scale applications
Facebook released Caffe2 but in March 2018, it was rolled into PyTorch
Allows users to switch from GPU to CPU by the setting of a single flag – first, it trains on GPU, then deploys on commodity servers
High-performance image processing that can process 60M images per day on one Nvidia K40 GPU
Provides model definitions, pre-trained weights, and optimization setting for quick starts
Features include denoising, depth estimation, optical flow, an semantic segmentation