Hacker News Reader

Comment by nextaccountic | Hacker News Reader

Does the Tensorflow model run in JS? How fast is that?

Would it be feasible to run the model in wasm or in the GPU?

I can't speak for this specific model/extension, but "most" neural networks aren't that big at the end of the day, and you are just a few matrix dot products away from getting your classification (of course, I'm drastically simplifying).

Its the training that takes forever, due to the fact all those numbers need tweaking. However one you have the model, classifying is pretty fast.

Are there tools to visualise trained model,?

Something like this: https://www.tensorflow.org/tensorboard

Something to visualise the models, the one in this link

https://github.com/nsfw-filter/nsfw-filter/tree/master/dist/...

Use netron to visualize model:

https://lutzroeder.github.io/netron/?url=https://raw.githubu...

This is amazing. Thanks for sharing! Didn't know about netron.

> Wait, it's all a DAG?

>> Always has been.

The brain is a heavily cyclic graph. We're not there yet ;)

I just installed this, and my whole browser grinds to a halt when I load a page...

That should not happen! Could you tell us about the issue? https://github.com/nsfw-filter/nsfw-filter/issues

Yes, tensorflow model can run on JS. I did that when developing a Deep Learning product, PredictSalary (https://predictsalary.com). The problem is not the model, but the TensorflowJS library (using Node) is big (8MB) after webpack-ing it (even after using compressing flag).

So I decided to move the inference to the server.

It's fast but I didn't toy with GPU setting when running Tensorflow model on JS.

So we use MobileNet V2 model. It is relatively light weight and can run fast on the browser itself

Reply on news.ycombinator.com