On February 16th, Google suddenly knocked on the blackboard at the microblogged public letter "Google Blackboard": the first TensorFlow Developer Summit was held in Mountain View, California, and officially released TensorFlow 1.0.
About TensorFlow, as early as 2015 we have reported that the United States local time on November 9, 2015 (Beijing time on November 10), Google Research announced the launchThe second generation machine learning system TensorFlow, and is open source.
TensorFlow is Google's internal machine learning system for many years, and its predecessor was the 2011 DistBelief internal machine learning program.
"Why should there be a real tool for researchers to try their crazy ideas.If those ideas have a role, then they will be able to convert directly into the product," he said at the time, "Why did we launch TensorFlow and open it?" There is no need for researchers to rewrite the code. "
TensorFlow is flexible, mobile, easy to use, fully open source and other features, anyone can use free of charge.
What are the new TensorFlow 1.0 improvements?
For the above characteristics, Google said, compared with the existing version, the launch of the new TensorFlow 1.0 mainly in the following improvements:
Faster:The TensorFlow team will release several model updates to show how to take full advantage of TensorFlow 1.0. It used the "beyond imagination" to describe;
More flexible:TensorFlow 1.0 introduces a high-level application interface with more modules for TensorFlow. At the same time, the TensorFlow team has announced the introduction of a tf.keras module that is fully compatible with Keras (another high-level neural network library)
Ready:TensorFlow 1.0 ensures the stability of the Python application interface. Python can easily get new functionality without breaking existing code.
In addition, TensorFlow 1.0 also has the following highlights:
1) The Python application interface is tuned to be more similar to NumPy;
2) Java and Go's experimental API;
3) After merging skflow and TF Slim, from the tf.contrib.learn and transferred to the high-level application interface module: tf.layers, tf.metrics and tf.losses;
4) An experimental version of the XLA (Specific Domain Compiler for TensorFlow Charts) was released for the central processor. XLA is developing rapidly, so it is expected that there will be more progress in the future conference;
5) the introduction of the TensorFlow debugger, which is TensorFlow run the command line interface and application interface;
6) object detection and positioning, as well as the camera image stylized new installation system demonstration;
7) Installation improvements: add Python 3 docker mirror and make the pip package compatible with PyPI. Now simply install "pip install tensorflow", you can install TensorFlow.
Google said TensorFlow ecosystems will continue to evolve by using dynamic batch technologies such as Fold, web tools such as Embedding Projector and updating existing tools such as TensorFlow Serving.
What are the practical uses of TensorFlow?
TensorFlow in November 10 last year (Beijing time) ushered in his first anniversary, in such a short period of time, it became the GitHub has the most forked repositories framework.
On this day, Google also carried out a treasure:
Track Australia's endangered manatees
Researchers use TensorFlow's latest image recognition technology to let the computer "learn to" identify the manatee in the giant aerial map, which is far more than the artificial, the accuracy is 1.4 times the naked eye.
Sorting and storing cucumber
A Japanese farmer used TensorFlow to build an automatic sorting storage system for the large number of cucumbers he had harvested. Automatically capture pictures, TensorFlow automatically sorted up to nine different quality levels.
Google Neural Network Machine Translation.
TensorFlow and Tensor Processing Units (TPUs) create a hardware accelerator for the Google Neural Machine Translate model. Translation error reduced by 55% to 85%. Very Google practice.
In addition, it is learned that TensorFlow is also used in more areas, including the diagnosis of Parkinson's disease, train positioning, writing music and so on.
Goldman Sachs in a 100-page report, the value of TensorFlow given a high rating:"TensorFlow has provided a precedent for leveraging company resources for other cloud platforms and research communities in terms of facilitating AI integration. At the same time, Google is leveraging its own strengths, such as the TPU, to leverage the open source world to compete for the company Advantage, although its machine learning library is open source.
On January 31, Google released TensorFlow 1.0.0-rc0, which runs the machine learning application on smartphone-level hardware.
Now, it finally ushered in the official version. In short, in the open source of this matter, Google has been whipped.