How is Biology Research handling new Tech?

>Submit paper on ML algorithm for spatial transcriptomics :marseysnappy:

>One reviewer insists we benchmark it against his two year old methodology with like 2 citations that no one knows or cares about :marseysmug2:

>Since scientists can't make repeatable or documented code the entire thing is hard coded to his one training dataset meaning I have to rewrite it :marseygigaretard:

>Rather then provide a GUIX archive or pack or a docker image they just have a list of pip packages :marseyboardcode:

>they are all so outdated they refuse to install the right version :marseytabletired:

:marseyeyelidpulling: I ofc have to finish this in 30 days while I also have two class projects due in 20 days plus studying for my finals. I love having to do the same work as PHD and post grad students for less pay :marseywagie:

TLDR; !linuxchads !fosstards !biology Despite having cowtools to create instantly deployable exact decency loaders via docker or GUIX scientists can apparently pass unreplicatable and hardcoded code past reviewers. :marseythumbsup:

Pip3 download speeds are prob butt (like 600 kb/s) because there is one gorillion AWS and Azure vms for ML trying to set up at the same time and amazon and Microsoft cant arsed to use their billions of dollars to make their own mirrors so they just leach off FOSS projects

52
Jump in the discussion.

No email address required.

Dogshit code is a part of the field. I do enjoy the silly design decisions you encounter when dealing with it.

Pip speeds are crappy because pypi was never intended for large wheels. Unfortunately the big ml packages shit all over the normal rules and have everyone downloading gigabyte Nvidia silo binaries for no good reason.

Jump in the discussion.

No email address required.

I have just never seen anything this hard coded :marseycringe:. Like as a reviewer you think question 1 would be "can this code be tested on other data". Like what is the point of publishing an ML model which works on only two datasets :marseyyikes:

Jump in the discussion.

No email address required.

It's pretty common. I often see models so overfitted that it's basically a top hat filter. "It works well on my data" is all these researchers understand

Jump in the discussion.

No email address required.

have everyone downloading gigabyte Nvidia silo binaries for no good reason.

I have to download this 4gb file to show a 15 second video of a black guy making a joke but it's somehow for "research". :marseyschizotwitch:

Jump in the discussion.

No email address required.

Link copied to clipboard
Action successful!
Error, please refresh the page and try again.