I don't really understand why people want to funnel themselves into these ml frameworks when you can have the freedom to implement arguably more interesting things in c++/fortran with libs like armadillo, openblas, superlu, etc? I mean, i get it if people just wanna copy paste the functions they see used in blogs/press/toy demo releases and get on with things, is that it?
I have an uneasy feeling about these frameworks as well, but I'm not an expert in this field. So if I'm told that it's not worth the effort to recreate all of that stuff from scratch I'm going to have to believe it.
What I do insist on is that we keep the capability to selectively dig deeper where we need to and combine different libraries with our own code.
That's why I'm always trying to reduce the frameworkness and increase the librariness of any third party code we use. Scripting language runtimes are a major roadblock in that approach.
> I have an uneasy feeling about these frameworks as well, but I'm not an expert in this field. So if I'm told that it's not worth the effort to recreate all of that stuff from scratch I'm going to have to believe it.
I wouldn't consider myself an expert in the "ml" field either, and only really informed because of my academic/contract research work in neuroimaging wrt DSP and beamforming and having to rely on performative code involving what at the end of the day comes down to manipulating matrices with some set of mathematical/physics constraints; I wished reviewers gave us an easier time with our paper and just believed our results without taking months before accepting it haha.
But if you are able to use these frameworks/hack around the quirks for your use case with minimum effort and solve your problems to satisfactory degree, I completely understand that they can be good enough.
>…we keep the capability to selectively dig deeper…
>That's why I'm always trying to reduce the frameworkness and increase the librariness of any third party code we use. Scripting language runtimes are a major roadblock in that approach.
I don't disagree with your sentiments at all, but it seems like you are already at the intersection where you just need to go down your own road because its probably unrealistic to expect Google's needs to be aligned with yours in the long term. Then again, I'm the type of person who would fork the nightly version of firefox and remove and add the stuff I want out of the box so maybe this is not the best suggestion >.<
