08:12 <discocaml> <bluddy5> if you're serious about ML, you have to use pytorch or tensorflow

08:12 <discocaml> <bluddy5> https://github.com/janestreet/torch

08:12 <discocaml> <bluddy5> seems decent but it's not compatible with pytorch 2.0 (not a huge deal).

08:12 <discocaml> <bluddy5> binding to python is definitely the easiest way to go.

08:13 <discocaml> <bluddy5> other solutions unfortunately aren't really relevant to state-of-the-art ML but may work for simple networks.

08:14 <discocaml> <bluddy5> Owl is good for manipulating arrays (which is a prerequisite for ML), but it lacks the ease of use and features of numpy, let alone pytorch's improved array-manipulation interface

16:59 <remexre> is there a hashtable with untyped keys and physical equality for keys?

17:07 <discocaml> <Kali> what would this be used for?

17:29 <remexre> I've got a bunch of nodes in a syntax tree that declare names, I'd be associating them with e.g. what scope they're in

17:37 <remexre> I suppose I could use Obj.t as the key, and equal = (==) with Hashtbl.Make?

17:59 <discocaml> <_korven_> I found this

17:59 <discocaml> <_korven_> https://github.com/LaurentMazare/ocaml-torch

17:59 <discocaml> <_korven_> In the README, it says you should continue using this for PyTorch 2.0.0

17:59 <discocaml> <_korven_> I found this

17:59 <discocaml> <_korven_> https://github.com/LaurentMazare/ocaml-torch

17:59 <discocaml> <_korven_> In the README, it says you should continue using this for PyTorch 2.0.0 support

17:59 <discocaml> <_korven_> whatchu say about this

18:10 <discocaml> <bluddy5> oh weird.

18:10 <discocaml> <bluddy5> yeah that looks fine then

18:10 <discocaml> <deepspacejohn> @remexre you could use a GADT to hide the type, like this: https://ocaml.org/play#code=bW9kdWxlIEtleSA9IHN0cnVjdAogIHR5cGUgdCA9IEtleTogJ2EgLT4gdAogIGxldCBoYXNoID0gSGFzaHRibC5oYXNoCiAgbGV0IGVxdWFsID0gKD09KQogIGxldCBtYWtlIHggPSBLZXkgeAplbmQKCm1vZHVsZSBVbnR5cGVkSGFzaHRibCA9IEhhc2h0YmwuTWFrZSAoS2V5KQoKbGV0IGtleV9pbnQgPSBLZXkubWFrZSAxCmxldCBrZXlfc3RyID0gS2V5Lm1ha2UgImEiCmxldCBoID0gVW50eXBlZEhhc2h0YmwuY3JlYXRlIDE2CmxldCAoKSA9CiAgVW50eXBlZEhhc2

18:11 <discocaml> <deepspacejohn> Since you use physical equality then you need to be careful about not creating a key more than once, since they would have to be boxed.

18:11 <discocaml> <bluddy5> I looked over the ocaml interface, and it seems pretty nice. the nice thing about Pytorch is that they reimplemented numpy, so you get that for free when you have bindings to it.

18:13 <discocaml> <bluddy5> The problem is that OCaml is missing things that exist in the python universe that make ML easier. You need easy graphing utilities, easy image display/print utilities. You *need* to be able to step through code in a debugger and examine variables.

18:14 <discocaml> <bluddy5> Libraries that do image augmentation. etc.

18:14 <discocaml> <bluddy5> Easier to bind to python.

21:57 <discocaml> <leonidr> 👋 Hello. Long term OCaml user getting a bit back into it. What's the state of the OCaml community these days? Where should one post stuff about confusing build errors in dune?

22:00 <discocaml> <._null._> Doing good ? You can put it in #beginners, #advanced-help or #dune depending on how advanced and specific it is

22:01 <Fardale> Or the forum at discuss.ocaml.org