This paper summarizes some features of tensorflow at Google cloud next conference held in San Francisco recently.
1. It is a powerful machine learning framework
Tensorflow is a machine learning framework based on data flow graph. It is the second generation machine learning system of Google brain. It is often used in various areas of deep machine learning, such as perception, language understanding, speech recognition, image recognition and so on. Tensors represent n-dimensional arrays and flows represent calculations based on data flow graphs.
If you have enough data and are in the stage of deep learning, neural network and advanced artificial intelligence of artificial intelligence, it may become your best helper. Tensorflow is not a tool, but a framework. If you want to return a regression line through a 20 × 2 spreadsheet, you can stop learning and start using it now.
When you want to achieve high-end and grand achievements, you will be excited. Tensorflow has been used in the space field to find new planets. It can help doctors screen diabetic retinopathy to prevent blindness, and it can also help save forests by warning illegal deforestation activities. Alphago and Google cloud vision are built on tensorflow, which you need to pay attention to. In addition, tensorflow is open source and can be downloaded for free and used at any time.
Kepler-90i, an extrasolar star discovered with the help of tensorflow, makes kepler-90 the only extrasolar Galaxy known to us, with eight planets orbiting it. No other galaxy has more than eight planets.
2. The method is optional
If you’ve tried tensorflow before, you’re scared out of it. Because it forces you to be like an academic researcher rather than a developer, but at present, there are more choices, so come back and use it.
Tensorflow eager execution allows you to interact with the system like a python programmer: all the real-time coding and debugging is performed on a per line basis, rather than writing large blocks of code like other languages is daunting. I am an academic researcher, but I like tensorflow eager execution from the beginning, so start using it as soon as possible.
3. Support line by line construction of neural network
Keras + tensorflow = easy to build neural network
Keras is a deep learning library based on tensorflow, which is written by pure python. It is user-friendly and can provide simple and rapid prototype design, which will be more helpful for some lower versions of tensorflow. If you like the object-oriented way of thinking, and prefer to build a layer of neural network at a time, then you will completely like tensorflow.keras. In the following lines of code, we create a coherent neural network with standard ringing tones and whistles that look like walking tones.
4. It’s not just Python
5. You can do anything in the browser
Tensorflow.js is used to evaluate the real-time human posture in the browser. Turn on your camera, and see this example.
6. Give a simple version to the micro device
Tensorflow Lite enables the model to execute in a variety of devices, on a single or multiple CPUs or GPUs of a personal computer or server, or even on mobile devices and the Internet of things (IOT). Tensorflow will give you more than three times of the original performance. It has good support for thread, queue and asynchronous computing. By making the most of the available hardware, you can freely allocate the computing elements in tensorflow data flow graph to different devices, and let tensorflow process the replica. Now you can start machine learning on a raspberry PI computer or mobile phone. In the speech at the conference, Lawrence did a brave thing, in front of thousands of people, through the image classification on an Android emulator, it really works well.
7. Professional equipment is better
If you’re tired of waiting for the CPU to complete the task of training the neural network with the data provided, you can now use cloud TPUs to specially design the hardware for this task. Just a few weeks ago, Google released the third generation TPU (tensor processing unit) on the alpha platform, an ASIC (integrated circuit chip technology) specially designed for machine learning and tensorflow. TPU is a programmable artificial intelligence accelerator, which provides high throughput and low precision computing (such as 8 bits), and is oriented to use or run models rather than training models.
8. Significant improvement of new data pipeline
Are you still using numpy in data pipeline? If you want to use tensorflow, tf.data namespace can make you more expressive in tensorflow input processing. Tf.data can provide you with a fast, flexible and easy-to-use data pipeline, as well as synchronous training.
9. You don’t need to start from scratch
We all know that machine learning from scratch is not an interesting way. Open the editor, there is only a blank new page, and there is no instance code. At this time, you can use tensorflow hub to continue an old habit of using other people’s code to help you write your own code, which is called your own code.
Read the original text
This is the original content of yunqi community, which can not be reproduced without permission.