Hacker News Reader

Comment by MrResearcher | Hacker News Reader

It's a pity that tfjs never truly developed any decent ops. E.g. you need lgamma to implement the cap for zero-inflated poisson regression and tfjs simply doesn't have that: https://github.com/tensorflow/tfjs/issues/2011 Those guys simply dropped the ball on the floor, and the further framework development fizzled out. I tried to find if WebGPU could be a replacement, but it's very hard to find a good description of the API surface. MDN also doesn't seem to help much – https://developer.mozilla.org/en-US/docs/Web/API/WebGPU_API I really hope that one day I'll find something similar to pytorch, only in typescript. Tell me I'll live long enough to enjoy that day.

View on HN

WebGPU does not play in the same league as tfjs, it is a successor to WebGL, the technology tfjs is implemented with. There is already a WebGPU backend for tfjs: https://www.npmjs.com/package/@tensorflow/tfjs-backend-webgp...

There are also several ONNX runtimes using WebGPU in case you want to run models written with PyTorch.

Exactly, the lack of the deep learning frameworks to train decent models on the client side forces us to use pytorch on the server, leading to centralized data collection, privacy concerns, copyright and data ownership issues.

Imagine the world where you could trivially train a model entirely on the client side, without uploading all that data to the cloud. Then we can settle on federated learning or simply use ensembles of models trained on different clients, all without sharing data with the server.

BTW, I did have some experience with ONNX, also ran into problems with some ops (like nn.SELU not working correctly in the browser - https://pytorch.org/docs/stable/generated/torch.nn.SELU.html...).

> Imagine the world where you could trivially train a model entirely on the client side, without uploading all that data to the cloud.

You just rediscovered "Federated Learning", but FL also allows you to benefit from the training being executed in other edge nodes, all without disclosing any data.

Instead of sending the data to the training server, they send the model to the users to train on a few batches of user data, then average the models from everyone.

It’s Hadoop all over again. Send the code to the data. To be clear, I think that’s an excellent idea and good to see it coming backs again.

Reply on news.ycombinator.com