90 comments

[ 3.3 ms ] story [ 172 ms ] thread
Seriously cool. Great work. I did get a glitch every now and then in the rendered output (say 1 out of 5 times) using Safari 10.1.2, MBP touchbar 2016 15", Radeon Pro 460 4096 MB.
Really cool demo. How does this compare to https://github.com/transcranial/keras-js ? Do the authors have a licence in mind?
TensorFire is up to an order of magnitude faster than keras-js because it doesn't have to shuffle data back and forth between the gpu and cpu. Also TensorFire can run on browsers and devices that don't support OES_TEXTURE_FLOAT.

We will probably release it under an MIT license.

I'm really interested in using Smartphones / Mobile devices for inference. Can this work with react-native so that I can build it without a bridge ? I would assume I would create a webview that would load a local website.
Where is the repo?
We're still finishing up a few things (documentation etc) and planning on releasing more stuff tomorrow.

You can also sign up for the mailing list if you'd like us to email you when the repo goes live!

Great. Look forward to diving in tomorrow. Thank you for the quick reply!
Would be great to port YOLO on your library; always an impressive visual demonstration
Hey HN!

We're really excited to finally share this with you all! This is the first of a series of demos that we're working to release this week, and we're hoping you'll keep us to that promise :)

Sorry if it doesn't work on your computer! There's still a few glitches and browser compatibility problems that we need to iron out, and we're collecting some telemetry data with LogRocket (https://logrocket.com/) to help us do so (so you all know what kind of data is being collected).

We'll open source the library under an MIT license once we finish writing up the API docs, and fixing these bugs.

This looks awesome!

It looks like it (like keras-js) is only for inference (running already-trained models) and not for training. Is this correct?

Are the operations or memory required for training very different?

Yes, you are correct! Training benefits much more from available memory through batching and, since in many cases you only need to train once, it usually makes sense to train on beefy GPUs.

TensorFire is useful in situations where you want to perform inference, but you don't want to ship user-supplied data to your servers, either because you would run out of bandwidth, you would run out of compute power, or your users want to keep their data private.

I'm not sure if it's working with my browser or not. It says "Compiling network", then shows a lot of flashing rectangles, then stops and displays a single grey rectangle. Is that what it's supposed to do?
Dude what it did was paint , as in recognizing thats its a picture of 2 cubs and then paint it like the way humans can, its freaking amazing
Just wanted to note, I ran the kitten demo in Chrome on my Nexus 6P (Android O Beta) and it worked perfectly.

Extremely impressed. Keep it up!

It's quite unreal. I remember when the paper and initial implementations came out less than 2 years ago, you had to go through this really long setup process that only worked on certain operating system and was a huge fuss. A few services came out online that would do it for you, but they were slow and limited, with huge queues.

Now, as you mention, you can run it in a few seconds on your phone, or in my case, on my Chromebook, right in the browser, with zero installation. Truly amazing.

It really had trouble with the portraits in my experience.
I quit working at shoprite and now I make $35h - $80h...how? I'm working online! My work didn't exactly make me happy so I decided to take a chance on something new…after 4 years it was so hard to quit my day job but now I couldn't be happier.

........... http://ow.ly/iBXm30dNtIZ

I quit working at shoprite and now I make $35h - $80h...how? I'm working online! My work didn't exactly make me happy so I decided to take a chance on something new…after 4 years it was so hard to quit my day job but now I couldn't be happier.

........... http://ow.ly/iBXm30dNtIZ

Very nice work, folks. Impressive, and very well-put-together demo. That's the easiest neural style transfer demo I've ever used - and the most fun. (Other than a minute of worrying that my poor 2013 MBP was about to melt down, but that's not your fault. :-)

The download link failed, as others have noted.

Thanks so much for sharing this!

Do you have benchmark number like FLOPS compared to GPU / CPU?
If I upload a 6MB image from my Canon, the site/browser (chrome) crashes. Example images work fine.
Works fine in Chrome on my Google Pixel, Android 7.1.2.
Very nice to see webgl gpgpu apps, they have been slow in coming. Any plans for webgl 2?
Kudos for providing a minimum experience on mobile! I was afraid I would have to wait until I got home :-)
Could someone explains whats is going on here? What are the steps? Why those colorful artifacts appear before the final result?
It´s just a demo of an upcoming open source API that allows running deep neural network models on the browser.

Steps (disclaimer: I´m not related to the creators, so this is just what I understand it does)

1.- You upload your image

2.- Select an image to be the origin of the style

3.- Downloading Model: downloads a trained (on style transfering) deep neural net

4.- Colorful artifacts: the model is applied to your image. Probably the artifacts are a visualization of the network weights being transformed to WebGL shaders, or just a simple visualization of the internal hidden steps of the transformation

5.- You get your image with the style applied

It's showing a visualization of all the intermediate activations of the style transfer network. The intermediate pictures are 4D, so they're visualized as a sequence of tiles.

The network being run is defined here https://github.com/lengstrom/fast-style-transfer/blob/master...

This post provides a pretty good explanation of what's happening: https://shafeentejani.github.io/2017-01-03/fast-style-transf...

There's a sequence of 9x9 and 3x3 convolutions that transforms that one big input image into a bunch of smaller images. They're processed by a sequence of residual convolutions. Finally, these tiny tiles are merged together back into a stylized image of the same size as the original input with a few deconvolution operations.

"running networks in the browser with TensorFire can be faster than running it natively with TensorFlow."

could you elaborate on this statement ?. What kinds of architectures does this hold true for ?.

From the github issue referenced in the FAQ, I think they mean that because TensorFlow only natively supports CUDA, TensorFire may outperform TensorFlow on computers that have non Nvidia GPUs, such as the new MacBook Pro.
Is the end goal to allow people to donate computing power for training? (a la Folding@home or SETI@home except just by visiting a webpage)

If so that's amazingly clever!

I think the goal is to allow people to develop webapps with models built using neural network libraries like keras and tensorflow. This would greatly improve the distribution of applications that are powered by deep learning because you wont have to install a bunch of dependencies in order to use the app.
I guess WebGL is now the "good enough" cross-platform vendor neutral replacement for CUDA.

Tensorflow should add a WebGL backend that runs in NodeJS.

Not quite. Training is not really supported in WebGL. For running a trained model this is cutting edge, and still has varying browser quirks.
Failed when I uploaded an image

>> framebuffer configuration not supported, status = undefined

This is amazing. I can't use GPU Tensorflow (natively) on my Macbook Pro because it doesn't have an NVIDIA graphics card. But I can... in the browser! Honestly didn't see that one coming.
To be clear, you can use it - just without GPU acceleration. The CPU-only build is supported and should work for you. If it's not, please let (me, us) know. Be sure to compile with AVX2 if you're on haswell or later; it helps quite a bit with some models.
> as fast as CPU TensorFlow on a desktop

> You can learn more about TensorFire and what makes it fast (spoiler: WebGL)

Does this mean that using a GPU in a browser through WebGL yields the same speed than a desktop CPU?

Actually it seems like WebGL is doing it even faster. Which makes sense - machine learning involves a lot of matrix math, which GPUs are made for and CPUs aren't.
This is actually what I'd expect, but the website feels quite misleading. Advertising that a GPU-based approach can outbeat a CPU for neural nets is not a very strong commercial claim :)
This is really cool! Great work!

I wanted to download the resulting image but got a "Failed - Network" error :(

>"Could not initialize WebGL, try another browser".

Happening in both Firefox and Chrome on Ubuntu. What exactly am I missing here?

For instance you might run on LLVMPipe or use some very old driver blacklisted in both browsers.

Firefox: about:support

Chrome: chrome://gpu

I've played around with doing some computation in WebGL, but it was rather tedious and difficult with my limited knowledge about the topic. It's possible, but you can't even rely on floating point texture to be available on all systems, especially mobile. And for anything more complicated, you probably need to be able to render to floating point textures, which is even more rare than support for plain floating point textures.

This only makes it more impressive when people do cool computational stuff in WebGL, but I'd wish there were some easier ways for non-experts in shader programming to do some calculations in WebGL.

WebGL 2 provides a much nicer base feature set and has been shipping & enabled in browsers since January or so.
TensorFire was a finalist of AI Grant. Applications for the next batch are open now! Get $2,500 to work on your AI project: https://aigrant.org.

It should only take five minutes or so to apply.

This is awesome!

Quick question: is the code compiled from js to webgl in browser as well, or do I need to compile beforehand?

I see this as a great way to learn and teach AI without having to bring a large toolchain.

Edit : it seems it is just a runtime for now for Tensorflow models!

Cool! Awesome! New MIT grads playing with JS who think "this whole deep learning thing is pretty neat"!

Most "normal" humans -- and application developers -- will just see this as "yet another webapp" running some crazy ass code in a browser tab with no apparent rhyme or reason.

We've gotta stop the insanity! We need regulation! We need: "Linux on the Web" (a web-based OS that I've been developing for the past 5 years, see it here: https://linuxontheweb.org) to start putting this kind of crazy, experimental stuff into actual use.

I'm really not trying to be too much of a jerk, but I've been working on my project for about as long as it takes to get an entire college degree, and I'm going to start needing some help here!

Respect. This pretty much killed the PC I'm on now. Wasn't even able to get to the task manager :D

Windows7, Firefox 54(64bit)

Nice, Leonid Afremov is a great choice of input art.