search

Psy-Fer / deeplexicon

Signal based nanopore RNA demultiplexing with convolutional neural networks
https://psy-fer.github.io/deeplexicon/
MIT License
36 stars 8 forks source link

Support for newer GPU architectures #25

Open lpryszcz opened 1 year ago

lpryszcz commented 1 year ago

Hi James, we are unable to build deeplexicon GPU version on newer GPUs (codename Ampere: RTX3080Ti, A4000). This is because those GPUs require newer CUDA version (deeplexicon is built against CUDA v10).

Could you please update the dependencies of deeplexicon to CUDA v11.8 so we can be build and executed it on the latest Nvidia cards?

lpryszcz commented 1 year ago

btw: interestingly, deeplexicon executes within container (Master of Pores) on newer GPUs, but the results are completely wrong. maybe we should warn other users about that?

Psy-Fer commented 1 year ago

Hey,

Yea I think I should probably just rebuild it with the latest tensorflow and re-validate the results. When we built it all, tf was still going through rapid changes along with cuda and various other libs. Things seem to have settled somewhat.

I'm currently waiting on some raw data from an RNA004 run to see what ont have done with the chemistry and how it impacts the signals, and if it would allow a better method of barcoding (that would be ideal)

What do you think?

lpryszcz commented 1 year ago

yeah, new chemistry will brake everything. can't wait :P

as for deeplexicon with CUDA v11, I guess we'll need it sooner than ONT releases new RNA chemistry. I'd love to play with it myself, but got tons of other things on my plate right now.

Psy-Fer commented 1 year ago

Yea I'm pretty slammed too. I was just waiting to see what kind of priority this was going to be with the new kit/flowcells. Should be a few weeks at most to get some data from Martin to get a somewhat early idea of what to expect. Shall we wait till then?

enovoa commented 10 months ago

Added a warning in the README on this issue