Source : "Dark knowledge - knowledge of how to subvert the machine business and society."
TensorFlow by Google brain development team, mainly for research and depth of machine learning neural network. May 2016, Google transferred from the Torch (a programming framework) to TensorFlow, which caused a blow to other programming framework, in particular torch and theanoo many people will TensorFlow described as a ratio theano more modern version, we learned all these years many important lessons in new areas / technologies.
TensorFlow intelligent, flexible way is known, it is a highly scalable machine learning system, making it easier to adapt to different old and new products and research, and relatively easy to install, but also provides a tutorial for beginners, covering neural network theoretical basis and practical application. TensorFlow than theano and torch slow, but Google and the open source community is addressing this problem. TensorBoard TensorFlow visualization module, which provides an intuitive view of a path computation. Deep learning library Keras TensorFlow been ported to run, which means that any model with Keras can now be written to run on TensorFlow. Finally, it is worth mentioning that TensorFlow can run on a variety of hardware, which is characterized as follows:
(1) GPU Acceleration: support
(2) the language / interface: Python, Numpy, C ++
(3) Platform: Cross-platform
(4) Maintenance by: Google
theano originated in 2007, well-known MILA (learning algorithms Institute) at the University of Montreal, is written in Python CPU / GPU symbolic expressions of deep learning compiler. theano powerful, fast, and flexible, but is generally considered to be a bottom frame. Therefore, native theano more like a research platform and ecosystem, rather than deep learning library, which is often used as the underlying platform for high-level libraries, and these advanced libraries to provide users with a simple API library includes some of the more popular Keras, Lasagne and Blocks. One of the drawbacks theano is still a need for multi-GPU support programs. Under theano features:
(1) GPU Acceleration: support
(2) the language / interface: Python, NumPYo
(3) platform: LinuxvMacosx and
(4) Maintainer: MILA Laboratory, University of Montreal
In all common framework, torch may be the easiest to get up and running, in particular, it allows the algorithm based on neural network running on the GPU hardware in the case of Ubuntu (an open-source computer operating system), without the need to hardware level encoded. torch developed in 2002 by the New York University, is widely used Facebook and Twitter and other large technology companies, and supported by NVIDIA. Torch is called in a scripting language, the language is easy to read, but not as common as Python. Useful error messages, a lot of sample code / tutorials, and Lua simplicity let torch is very easy to use. Its characteristics are as follows:
(1) GPU Acceleration: support
(2) the language / interface: Lua
(3) platform: Linux, Android, Mac OS X , iOS and Windows
(4) defenders: Ronan, Clément, Koray and Soumith
Caffe been developed for image classification convolution neural networks / machine vision by more than 1,000 developers to promote their development. Caffe perhaps best known is Mode1Zoo model, developers do not need to write any code can be used directly.
Caffe mainly for industrial applications, while the torch and theano is tailored for the study. Caffe apply to non-visual depth study computer applications, text, sound, or as time-series data. Caffe can run on a variety of hardware, and switching between the CPU and GPU may be done by setting a single flag. Caffe runs faster than theano and torch slowly. Its characteristics are as follows:
(1) GPU Acceleration: support
(2) the language / interface: C, C ++, Python, MATLAB, CLI
(3) platform: Ubuntu, MacOSX, windows experimental version
(4) Maintainer: Berkeley Vision and Learning Center (BVLC)
CNTK is Microsoft's deep learning toolkit, Microsoft's open source depth learning framework. CNTK community in general is more famous deep learning in the community than in voice, image and text can be used for training. CNTK support a variety of algorithms, such as Feed Forward, CNN, RNN, LSTM and Sequence-to-Sequence. It can run on many different hardware types, including multiple GPU. Its characteristics are as follows:
(1) GPU Acceleration: support - [
(2) the language / interface: Python, C ++, C # and CLI
(3) platforms: Windows, Linux
(4) Maintainer: Microsoft Research
H20 also called H20.ai, it is one of the world's most widely used open source depth learning platform. It is the world's more than 80,000 data scientists and researchers, and more than 9000 companies and organizations use, including the development of mission-critical data products for the world's most influential companies. H20-based user interface, and you can access the machine learning software libraries, and open the process of machine learning.
Wikipedia has a detailed table lists the parameters and characteristics of each major programming framework link below: https: //en.wikipedia.org/wiki/Comparison_of_deep_learning_software.