This article appeared in the Xinzhi yuan (micro signal: AI_era) author: Simon
In April 14, 2016, Google released a distributed TensorFlow. Google's blog introduced the TensorFlow in the image classification task, in the 100 GPUs and less than 65 hours of training time, to achieve the correct rate of 78%. In the fierce business competition, the faster training speed is the core competitiveness of the artificial intelligence enterprise. And distributed TensorFlow means that it can be truly large-scale into the artificial intelligence industry, have a real impact.
Google yesterday released a distributed version of TensorFlow!
According to Github's statistics, TensorFlow has become one of the six most popular open source projects in 2015. Taking into account the TensorFlow only released in December, a month's time to let it become the focus of the world's attention.
But at the time of the TensorFlow, is only a single machine running on a single version. This means that although the design is exquisite, but it is difficult to be used by companies, organizations, it is difficult to cause a substantial impact on the industry.
But the distributed TensorFlow released yesterday, the most prominent feature is to be able to run on different machines at the same time. Although not all people need to run TensorFlow on several thousand servers, but the researchers and Venture Company are indeed able to benefit from the TensorFlow of multiple machines.
After 5 months of waiting, the distributed TensorFlow finally arrived.
TennsorFlow 0.8 is released, and it has some good improvements. It makes some changes to the distributed version, and it makes them easier to use. This blog also describes the use of distributed systems to train a number of scalable digital image recognition model.
Google official blog introduction
TensorFlow is an open source software library developed for the numerical computation of data flow diagrams. The nodes in the graph represent mathematical operations, while the edges of the graph represent the multidimensional array of data (Tensors). In the case of using only a single API, a flexible architecture allows you to deploy a single or multiple CPUs and GPUs deployment on a desktop, server, or mobile device. TensorFlow the earliest by Google brain team of researchers and engineers to management of machine learning, neural networks, and depth of the research work, but the system is generic enough to be applicable to other application areas.
TensorFlow 0.8: support for distributed computing
Google in a lot of products are used in machine learning technology. In order to continuously improve our model, the most important is to train speed as fast as possible. To do this, including a way is in the hundreds of machines running TensorFlow. This can put part of the model training process from weeks reduced to a few hours, and can let us in the face of the increasing size and complexity of the model, and also to carry out the experiment. Since we open source TensorFlow, a distributed version becomes one of the most needed features. Now, you don't need to wait any longer.
Today (April 14th), we are excited to introduce the TensorFlow 0.8, which has distributed computing support, including all support for training distributed models on your infrastructure. Distributed gRPC supported by high performance TensorFlow libraries, but also to support parallel training on several hundred machines. It complements our recently announced Google cloud machine learning, but also be able to use the Google cloud platform to train and serve your TensorFlow model.
Figure: TensorFlow can accelerate the speed of training to generate the network, the use of 100 GPUs can reach 56 times. Source: Google
Distributed training also supports the use of cluster management systems such as Kubernetes to expand the size of the training. Further, once you have trained the model, you can deploy to the product and accelerate the speed of reasoning in TensorFlow using Kubernetes services.
At present, the TensorFlow version of distributed computing is just a start. We will continue to study ways to improve the performance of distributed training, both through the project, but also through the improvement of the algorithm, we will also share these improvements in the GitHub community.
If you want to skip the complex process, feel Google, TensorFlow provides a browser based simulator that allows you to experience the basic TensorFlow and deep learning.
First, select the data you want to analyze on the left, and then select and combine the items in the middle, and finally see how the output is matched to the earliest data. At first it seems ridiculous, but it's easy to understand and understand how the neural network works in an abstract level.
TensorFlow update history
Learning in the Cloud with, TensorFlow Machine
Neural network image classification Scaling using Kubernetes with TensorFlow Serving
Your own image classifier Train with Inception in TensorFlow
Your models in production with TensorFlow Serving Running
Yourself Deep Learning with TensorFlow and Udacity Teach
To Classify Images with TensorFlow How