Hacker News Reader

Comment by zawerf | Hacker News Reader

One very practical reason is to save on computation costs. GPU servers are not cheap and if you're going to run on CPU anyway, you might as well run it locally on the user's browser.

Some demos such as real time object detection isn't possible at all if you had to pay roundtrip server latency (not to mention the complexity of streaming video to and from your server): https://github.com/ModelDepot/tfjs-yolo-tiny

And a lot of web demos such https://affinelayer.com/pixsrv/ I don't believe will be up for such a long time if the author had to pay ongoing server costs.

View on HN

There is NO save on computation cost. CPU is slow for NN whatever language you are using.

OP has a point, if I want to learn PRACTICAL front end programming, I will choose javascript, not Python. Same for Neural networks, just switch the place.

What's wrong with js for neural networks? JS/V8 is faster than python, not accounting for python bindings to native NN libs that may as well be bound to js.

There is nothing wrong to use js, like there is nothing wrong to call it impractical.

JS is faster than Python, but in the land of DL, there is C++ and everyone else. No one is using Python to do the actual computation anyway.

"No one is using Python to do the actual computation anyway."

Depends what "actual" computation is. If you're definition of "actual computation" is something which requires extreme optimization then your definition precludes the question. In the real world there is an extraordinary amount of computation done with inefficient languages though simply because development time costs very often outweighs run time costs.

Python is used because there are C/C++ bindings for working with NNs.

The point is that none of computations are done in pure Python. Python just provides a convenient wrapper over non-Python code.

JS/V8 may beat Python in general operations but definitely not in math specialized libraries. A lot of python packages support OpenBLAS and AVX/FMA by default which gives huge boosts to math/tensor operations. Correct me if I'm wrong, I don't think the js ecosystem is quite there yet except for experimental projects.

Cost as in money.

To deploy a model in js (as a web page) all you need is a static S3 or GCS bucket. You don't even need a webserver and it can automatically handle infinite scale. Show me a python solution that can do the same.

There is currently no NN learning algorithm that can handle massively paralleled training. We can use some simple fixes such as mean gradient but they have severe limitations, and their limitation grows as you scale up. Currently even with a handful paralleled GPU training, the gradient computation needs to wait for all GPU batch to complete and then return to CPU before the next batch, so your idea of infinite scaling is just a pipe dream at the moment. Sure Python can't be run on a million pages, but heck, no NN architecture can even handle more than a dozen paralleled computation in a stable manner yet.

Yea the infinite scaling was referring to deployment of inference, not training.

> To deploy a model in js (as a web page) all you need is a static S3 or GCS bucket.

First of all, what you have described is far from the reality.

Had this come true, only inference will be in javascript, through some language agnostic standardized model format, not training. The model is just a blackbox function for the js runtime to call. The amount of javascript to make this happen will be surprisingly slim anyway.

> First of all, what you have described is far from the reality.

Did you try any of the links I included? This is the reality for all of them and they are a few years old. They have a model file loaded from bucket url and never make another network request thereafter.

It's actually the recommended workflow from https://js.tensorflow.org/ where you'd find tons of other examples.

But you're right, this is for inference only. I would not do training in JS.

What I mean reality here is the running nn model in browser is possible, but not practically efficient, so even with its perceived simplicity, people would under most occasions, run the inference in the cloud, with its controllability and performance, like using customized hardware. After all, running inference is about to run it reliably and fast, until the day when nn operations are ubiquitous and common enough to be standardized and shipped in performant runtime that come out-of-box, browser inference is still a dream that is too good to be true.

Are you suggesting distributed NN training through user's browsers? This is a very sketchy area along the lines of browser based bitcoin mining. You don't want to run any computation heavy code on client side without user's knowledge & approval.

Though if you only meant performing inference in user's browsers, then the challenge would be to find a way for TF/pytorch pre-trained models to port and perform accurately in js. If successfully done, I can see some use cases here.

See https://js.tensorflow.org/ for automatically porting tensorflow models to the browser

Are you under the impression that js is the only language that can run locally?

in a client's web browser... yeah.

Well there is web assembly, but why is running in a web browser a requirement for creating a neural network anyway?

so you can offset server costs to the client which usually means bloated laggy apps until someone comes along and figures out an optimised way of doing things, for eg react.js vs phonegap's comparative shitshow

Reply on news.ycombinator.com