This morning, I found out that I was attending the "Google Cloud Next" conference in San Francisco, USA, where I enjoyed the lectures and summarized the favorite points. The theme of this lecture is "What's new in TensorFlow?(YouTube video)"
Then, I thought about it for a while, and thought about not sharing my super short summary with you (except maybe you won't read this video, but you should really look at it because the speaker is great). So let's get started!
1.TensorFlow is a powerful machine learning framework
TensorFlow is a machine learning framework. If you have a lot of information and/or you are pursuing the most advanced technologies of AI artificial intelligence, such as deep learning, neural networks, etc., then TensorFlow may become your new super friend. TensorFlow is not like the Swiss knife of data science. Its function is very big. It is comparable to industrial lathe. I mean, if all you want is to draw the regression line through 20's 2 spreadsheet, then you You don't have to continue reading.
But if the huge utility is what you are after, then TensorFlow will make you feel excited! Because it has been used to find new planets; to help doctors screen for retinopathy caused by diabetes to prevent blindness; to remind the authorities to illegally deforestation signs to save the forest. AlphaGo and Google Cloud Vision are built on top of TensorFlow, and now you can try it out! TensorFlow is completely open source, you canfree downloadAnd start using it right away.
With the help of TensorFlow, scientists discovered the Kepler-90i planet. This makes the Kepler-90 galaxy the only other galaxy known so far, with a total of 8 planets in orbit around a single star. So far, no other galaxy has been discovered with more than 8 planets, so I guess that our solar system and the Kepler-90 galaxy are temporarily at the top of the list! To learn more, click here.
2. No need to write programs in strange ways
I fell in love with TensorFlow Eager.
If you have used TensorFlow in the past, and because it forces you to write a program like a scholar or an alien, rather than a professional programmer, screaming away. Remember to hurry ~ fast ~ back ~ come ~ 喔!
TensorFlow's eager execution lets you interact with TensorFlow just like a pure Python engineer. Being able to write programs and debug directly, one by one, smoothly, instead of holding your breath, carefully constructing those huge maps. I have returned to my university as a lecturer (and probably an alien) in the past few years, but since the release of TensorFlow's eager execution, I have fallen in love with it!
3. You can build neural networks one after the other.
Keras + TensorFlow = Easier to build neural networks!
HardSince its release, it has been loved by users because of its user affinity and ease of prototyping. These two points have always been the old version of TensorFlow, which is very eager to have. If you like object-oriented thinking and love to build neural networks layer by layer, you will fall in love with tf.keras. In the image below, with just a few strokes, we have created a sequence of neural networks that include standard accessories like dropout. (Remember to remind me that I want to talk about the dropout metaphor, they include the stapler and the flu.)
Hey, you like puzzles, right? Be patient, don't always think about the stapler!
4. No longer only Python
5. Users can train and execute models in the browser
6. For small devices, TensorFlow has a lightweight version.
Get a lost desktop computer from the museum? toaster? (The same thing?)TensorFlow LiteThe model can be executed on a variety of devices, including mobile devices, IoT devices, etc., in terms of speed of inference, more than three times faster than the original TensorFlow. Yes, you can now do machine learning on the Raspberry Pi or on your phone. In this paragraphLecture(YouTube video), the speaker Lawrence did a brave thing, is in front of thousands of people, on the Android simulator, live display image classification, and ⋯⋯ success!
7. Special hardware performance is better
If you are tired of waiting for the CPU to process the data to train your neural network, you can now get the Cloud TPUHardware designed to handle large amounts of data. The T of TPU means "Tensor", which has the same meaning as Tensor of TensorFlow. Coincidence? I don't think so! A few weeks ago, Google announced the third version of the TPU, which is currently in development to the preview version.
8. Significant improvement in new data workflow
You are usingNumpyWhat to do? Suppose you want to use numpy in TensorFlow and then quit immediately, nowTF.data namespaceMake your input processing in TensorFlow easier to express and more efficient. Tf.data for youFast, flexible and easy to use(YouTube video) data workflow and synchronization with training.
9. You don't have to start from scratch
Do you know a boring way to start machine learning? That is to display a completely blank new page in your editor, and there is no sample code at all. Now, with the help of TensorFlow Hub, you can join a long-established tradition of improved, more efficient, the tradition of accessing someone's code and calling it your own (also known as "career" Software Engineering").
TensorFlow HubIt is a library of reusable, pre-trained machine learning model components that are packaged for single-line reuse. So come on your own!
This is a summary of my summary. So, here's the full lecture, and enjoy the next 42 minutes!
This article is reproduced from the medium,Original address(Requires Internet Science)