"TensorFlow.js" appears that enables machine learning to be executed in the browser

Google updated the open source machine learning library "TensorFlow" at TensorFlow Dev Summit 2018 held on March 30, 2018, and "TensorFlow.jsAnd "TensorFlow Lite" which can run on a small power machine such as smartphone and Raspberry Pi.

Introducing TensorFlow.js: Machine Learning in Javascript
https://medium.com/tensorflow/introducing-tensorflow-js-machine-learning-in-javascript-bf3eab376db

The state of TensorFlow Developer Summit 2018 can be seen from the following. There are about 8 hours in full story.

TensorFlow Dev Summit 2018 - Livestream - YouTube


Several demonstration programs using TensorFlow.js are released. The first one is to use the web camera to operate the PAC-MAN. actuallyDemo pageAlthough it was a thing that could be tried with, at the time of article creation it was not possible to ascertain the operation without progressing from "Loading mobilenet ...". If the code isPublished on GitHubIt has been.


Another demo isFind pictograms instructed using smartphone cameras in the real worldsomething like. Although it seems that it may not work depending on the environment, it seems to work when accessing with Safari on iPhone. me tooSource codeIt has been published and it is quite helpful.


Also, in TensorFlow.js, in addition to creating models directly on the browser, by importing already trained models, it is possible to train models quickly with less data. The actual wayTutorial pageAn example is stated in, and it is possible to start using easily. Looking at the schematic diagram of API of TensorFlow.js, in addition to the high level API for defining models, linear algebra andAutomatic differentiationThere is also a low-level API for, so you can see that both the Keras model and TensorFlow SavedModel can be imported.


It is an optimal library for people who "have experience of Javascript but machine learning is a novice".