Mnist Demo Compiled On-the-Fly to WebAssembly in the Browser | Hacker News Reader

Mnist Demo Compiled On-the-Fly to WebAssembly in the Browser

mnist-demo.glitch.me

52 pointsbrrrrrm21 comments

Is this supposed to recognize a single digit? For me it can recognize only half of the digits correctly. What is the use case for sich a low-precision image recognition?

Judging by the fact the page gives no information about the model used but says it's "compiled on-the-fly to WebAssembly by loop_tool.js", I assume it's intended as a demo of loop_tool.js, not the model. It can't be a demo of MNIST itself. That's just a dataset, presumably used to train this model.

roll your own i-am-not-a-robot capture

It is trained on the mnist dataset, to have it recognize a digit you have to make sure the digit is sized roughly to fill the entire screen. It won't work if you draw a small digit in the corner. This is just a demo of the toolchain, it doesn't showcase any state of the art image recognition.

Ah, so it's a toolchain demo. That makes more sense, and I was also initially confused; it kept saying my 5s looked like 8s most of the time lol

Wow the source code is kind of blowing my mind. I had no idea this was possible (only in theory).

  import * as torch from "./torch.mjs";

  ...

  const w0 = torch.tensor(weights["conv1.0.weight"]);
  const b0 = torch.tensor(weights["conv1.0.bias"]);
  const w1 = torch.tensor(weights["conv2.0.weight"]);
  const b1 = torch.tensor(weights["conv2.0.bias"]);
  const lw = torch.tensor(weights["out.weight"]);
  const lb = torch.tensor(weights["out.bias"]);

  const convs = nn.Sequential(
    nn.Conv2d(w0, b0, 1, 2),
    nn.MaxPool2d(2),
    nn.ReLU(),
    nn.Conv2d(w1, b1, 1, 2),
    nn.MaxPool2d(2),
    nn.ReLU()
  );
  ...

Looks like near identical python torch code except apparantely running in local js!

this is actually a brand new proposal! If you like it or have any comments, you can find more information here: https://dev-discuss.pytorch.org/t/proposal-torch-js-a-javasc...

The quality is, ehm, not really good: https://imgur.com/a/uobGnpK

Agreed https://imgur.com/a/m1DnaXP

Cool demo nontheless!

The same thing happened to me: https://i.imgur.com/yN9JhGw.png

Quite something indeed: https://imgur.com/a/SgqvG75

Same experience: https://i.imgur.com/Z9BrQGn.png

I don't hold the in browser POC at fault for the trained model (but am amused by the efficacy of the model).

Unrelated to this demo, does anyone have pointers or recommendations for deploying 3D/WebGL + Wasm to the web?

Any good libraries, tools, or resources?

You can find a working example in my Torus Trooper project: https://github.com/speps/tt

In particular, the interface between WASM and WebGL: https://github.com/speps/tt/blob/master/web/tt.js#L45

WASM is really quite straightforward in my opinion, export a C interface and link it with whatever you need from the browser APIs.

This is a fantastic reference, thank you! Cool project, too.

Cool idea, I wonder if python wasm doesn’t make more sense though

The startup was excruciatingly slow ~20s, is it mining crypto in the background?

It is "compiled on the fly", including (hardware specific?) optimizations, which can take some time.

That's interesting. It was nearly instant on my iphone.

Would you mind sharing which device you're using? I'm still working on making the first compilation faster!

The code is compiled with Ecmascripten and the WASM binary is injected as part of the JS module. Since the downloading, processing and compiling is happening on the browser, it seems to be a reasonable time.

However, to avoid blocking the main thread, I would run all this process in a worker. Not sure if there's any limitation there

Reply on news.ycombinator.com