Sender | Message | Time |
---|---|---|
23 Aug 2022 | ||
In reply to @jboy:utwente.ioCould you share an example of such a shell.nix? I'm interested in how to use a venvHook. | 14:29:17 | |
In reply to @carlthome:matrix.org I've been dabbling with jupyterWith but feeling that poetry2nix is easier when one just needs to get stuff done. But I've only tried poetry2nix for making a library so far, not a notebook environment: https://github.com/carlthome/poetry2nix-test/blob/main/flake.nix | 14:31:08 | |
| 14:31:57 | |
*
| 14:32:05 | |
*
| 14:32:26 | |
In reply to @jboy:utwente.ioThanks! | 14:35:02 | |
jboy: Sounds like a pragmatic solution if packaging is out of reach. Thanks for the snippet | 14:35:11 | |
carlthome: Have you also a nix file that uses jupyterWith? Where do you see friction with jupyterWith? | 14:36:08 | |
And thanks for the poetry2nix file | 14:36:19 | |
In reply to @flxai:matrix.org TBH not sure what I'm doing really. 🫡 Figured it could be convenient to have a set of regular ipykernels (one for PyTorch, one for scikit-learn, etc.), ready for playing around with ML modelling ideas. Thus I started just throwing in them as devShells in my ongoing home flake to learn more about jupyterWith. Here's a wip commit: https://github.com/carlthome/dotfiles/pull/3/files#diff-206b9ce276ab5971a2489d75eb1b12999d4bf3843b7988cbe8d687cfde61dea0R57 | 14:43:54 | |
Yes I feel you. There are many solutions to this problem. Will try those posted. Thanks again | 14:54:50 | |
24 Aug 2022 | ||
19:15:39 | ||
23:00:44 | ||
25 Aug 2022 | ||
17:46:09 | ||
30 Aug 2022 | ||
13:42:27 | ||
14:42:54 | ||
14:56:16 | ||
16:51:54 | ||
5 Sep 2022 | ||
05:39:36 | ||
18:21:51 | ||
6 Sep 2022 | ||
Nix is attractive for reproducibility and provenance tracking. However, for ML workflows one major limitation is that all data produced as part of a derivation resides in /nix/store. It would be great to be able to define flakes for large datasets and use them as inputs to other programs/flakes. For example define a flake for ImageNet and then get train and test splits as outputs of those downstream flakes. Currently such a train/test split would require duplication of the original dataset twice into /nix/store if I'm not mistaken. | 15:34:39 | |
apologies if there is an existing solution to this kind of data handling with nix that I'm unaware of | 15:35:12 | |
i guess the splitting could of course be done with symlinks. main question is then is there any hope of handling large datasets at rest on some storage outside /nix | 15:39:06 | |
Nix master supports impure derivations. Impure derivations have network access, so you could write your result somewhere else. I have not tried this yet, but I imagine this going to be very useful also in CI/CD. | 17:04:49 | |
https://github.com/NixOS/nix/pull/6227 | 17:05:19 | |
thanks. I currently use --impure on non-NixOS with nixGL to get cuda working anyway, so that may be the ticket | 17:07:24 | |
that's not the same though, that's impure evaluation | 17:08:40 | |
i see that now. should've read before replying.. | 17:09:23 | |
7 Sep 2022 | ||
Has anyone got a working example of
...but it keeps segfaulting later when I prompt What helps is running the | 12:43:40 | |
15 Sep 2022 | ||
12:10:46 |