Real-time semantic segmentation in the browser - Made with TensorFlow.js

[@anabakhtar3774]

Tensorflow JS is literally limitless 💯❤

[@hugoznn]

Yes!! Let's bring machine learning to the browser!

[@hiankun]

Same background here (CV+Python). I also found that running apps in browsers is such a powerful and attractive option. Thanks for such a bright mood talk. :-)

[@edgarssensalo7567]

Wow this is great. Python together with JS in ML will be unstoppable

[@sultanakalsekar6800]

Absolutely Correct ...

[@hugoznn]

Yes! A whole world of new opportunities....

[@paedrufernando2351]

@2:26 is where it starts

[@superaluis]

Is this running on some GPU or its purely on CPU? If it's running in a browser on CPU so can It be run in real time like that on a mobile phone?

[@JasonMayes]

TensorFlowJS can run on GPU in browser via WebGL

[@rodmallen9041]

Fantastic and extremely useful. Tks a lot!

[@jay7670]

Hello I am js developer and I want to start ML and I find out this js library on Internet so any one can please guide me how to start it and where to learn and what type of device I needed ?

[@hugoznn]

Hey, I recommend having a look at the tensorflow.js tutorials (www.tensorflow.org/js/tutorials) and demos (www.tensorflow.org/js/demos). You do not need any special device.

[@JasonMayes]

Hello there, to use TensorFlow.js you just need any computer or device that can run a web browser like Chrome / Firefox etc or Node.js (Raspberry Pi, or a cloud server for example). You can get started on the TensorFlow.js website. I suggest you check out our premade models as they are very easy to use. Also try Teachable Machine to recognize a custom object in under 5 minutes using JavaScript in the web browser!

[@JasonMayes]

PS the links are here: www.tensorflow.org/js teachablemachine.withgoogle.com/train

[@jay7670]

@@JasonMayes thank you brother I will start 🙂

[@DevsLikeUs]

Awesome! Thank you for sharing.

[@hugoznn]

Thanks for watching :)

[@Ceelvain]

It's pretty weird to hear someone who want something light and fast to turn to JavaScript. If you want something fast, the first and most important step is to get rid of the browser. Not to add one.

[@dimitrioskyriakos3243]

Use C++

[@JasonMayes]

Both Python + Node.js implementation of TensorFlow talk to the C++ bindings behind the scenes so for server side its essentially C++ under the hood. In the browser your only option is JS + WASM if you want the reach and scale of the web for folk to be able to use your model at scale without installing linux, installing tensorflow, installing cuda, cloning your git repo, reading your read me, and if you haven't lost the person by then hope it works and nothing has changed for the deps. :-) Of course you could host server side but then you have latency issues for real time applications, not to mention the cost of running servers for inference for a reasonably popular site can be very high so is not everyone's cup of tea.

[@dimitrioskyriakos3243]

under the hood is working binary code (machine language!) all thease C++ stuff are compiled to native code that means that's speaking the machine language. The scripting languages are not compiled every time that the interpreters and the runtimes are reading a scripting command it needs sometime to translate this command to a list of machine language code. That needs more time for execution. An other thing that I'm speaking about compiled languages when we have to manage heavy processes is the optimization. option 1 optimization trough the compiler options. option 2: optimization by disassembly the code and optimize the assembly code and after using the assembler to check about reusable code to put it in a routine and call it from there for example or the opposite depend on execution time. So for example you can create a server side a binary code (with C++, object pascal etc) that it can commuicate with your main app using A.P.I. or sockets. For me is absolutely wrong to execute Machine Learning in Fronded(on browser) that it will need resources from user's computer. Sorry for this big text.

[@JasonMayes]

@@dimitrioskyriakos3243 our Node.js server side implementation actually calls C++ under the hood just like Python does. But you can't do this in the browser which is why it is the way it is, and that is good enough for many usecase that previously were not possible. We can not always execute on server side for privacy reasons / regulations etc.