Hacker News Reader

Comment by sdan | Hacker News Reader

I've tried many of these papers, and they don't work outside their datasets.

Suppose you tried this on your own camera or maybe in a different lighting: it would totally break this.

Paper definitely seems like an improvement, but I guess no one should get excited to use this in experimentation/production.

View on HN

I've tried a number of ML projects like this too, and sometimes they do work – Bodypix is one example (though not precisely the same thing).

That being said, its usually pretty difficult to get code from papers to build/run successfully due to dependencies (e.g. depending on a specific version of Python and OpenCV 2 and requiring CUDA support)

Could anyone explain why version numbers make ML stuff so brittle? Why is the CUDA version supposed to match? Why would a newer version of python (3.7 vs 3.6) would ever break anything?

TensorFlow is a crapshoot as far as I understand, constantly changing making newer versions incompatible; but why do other libraries break backwards compatibility without a major version bump?

Because software is hard.

You might be relying on a function that only exists in 3.7 and not in 3.6, code written in 3.6 would work but new code using 3.7 features won’t be backwards compatible. With compiled code the errors are usually very hard for “more used to scripts“ people to decode. You get stuff like missing symbols in the linker phase.

ML projects usually have a lot of libraries so you also get in the transient dependencies breaking quite often...

Yeah, managing dependencies is tricky. I've been super excited about Nix and Guix lately for that reason; if you have a single Nix/Guix revision and a list of packages, you have all the information you need to build exactly the same package tree. (With bit-for-bit reproducibility where possible, no less!)

Some language-specific package managers can do similar things, but you really only get reproducibility for the whole system with a general-purpose package manager. Poetry gets you pretty far within the Python ecosystem, but if you need a specific version of Python/specific native libraries/etc... it doesn't get you all the way there.

Because dependency management for anything C/C++/Python related has often been a massive cluster* * * *. Rust's npm style package management is probably the greatest innovation to have hit low level programming in a long time. Not to mention anything Nvidia related is often closed sourced so you are programming against a black box API by a vendor infamous for their bad driver quality on non-Windows systems.

I tried to see what would happen when I put in some stock photos of people into the Colab notebook linked from the repo's README. Some worked better than others. Some totally didn't resolve well at all. Overall I think it's interesting, but there are definitely a lot of edge cases.

Some gifs of the test results: https://github.com/cardboard-q/pifuhd_demo_model_test

"yoga" example is quite disturbing.

For what it’s worth, this project did get an update about a month ago. I can think of some cool uses for this if it works.

https://colab.research.google.com/drive/1GFSsqP2BWz4gtq0e-nk...

Interesting cool applications from when I was a PhD student [1,2].

[1] http://www.eurecom.fr/en/publication/3189?&theme=mobieurecom

[2] http://www.eurecom.fr/en/publication/3247/copyright?popup=1

Reply on news.ycombinator.com