提问 HN:Tensorflow.js 凉了吗?
3 分•作者: fouronnes3•5 个月前
从提交活动来看,自 2023 年底以来,它似乎已经停滞不前,新提出的问题也基本被忽略。 它是被放弃了,还是仅仅进入了“成熟期”? 文档中仍然有很多关于“未来工作”的提及,所以我认为它已经被放弃了。
这很可惜,因为它仍然是唯一一个能提供以下良好实现的库:
* 浏览器中的数值 N 维张量
* 多个后端(cpu、webgpu、wasm、webgl)
* 自动微分
所有从事 LLM 和其他 ML 应用的人都只在浏览器中进行推理,因此他们对从其他框架导出的 ONNX 感到满意。但这并不能涵盖您希望在浏览器中进行优化并在 JS/TS 中创建自定义模型的情况。我甚至不需要所有的层和神经网络相关的东西。 只需要一个干净的张量 + 自动微分库,并支持 TypeScript。
查看原文
Looking at the commit activity it seems pretty much dead since late 2023, and new issues are largely ignored. Is it abandoned or just "mature"? There's still a lot of "future work" references in the doc, so I think it's just abandoned.<p>It's a shame because it's still the only library out there that provides a good implementation of:<p>* numerical N-dimensional tensors in the browser<p>* multiple backends (cpu, webgpu, wasm, webgl)<p>* autograd<p>Everyone doing LLM and other ML applications is doing inference only in the browser, and so they are happy with ONNX exports from other frameworks. But this doesn't cover the use case where you want to do optimization in the browser and create custom models in JS/TS. I don't even need all the layer and NN stuff. Just a clean tensor+autograd library with typescript support.