Newest at the top
| 2025-02-25 11:23:55 +0100 | <ski> | [exa] : not afaik |
| 2025-02-25 11:23:36 +0100 | <lambdabot> | ST s a -> a |
| 2025-02-25 11:23:35 +0100 | <ski> | @type Control.Monad.Primitive.unsafeInlineST |
| 2025-02-25 11:23:33 +0100 | <lambdabot> | ST s c -> c |
| 2025-02-25 11:23:32 +0100 | <ski> | @type System.IO.Unsafe.unsafePerformIO . Control.Monad.ST.Unsafe.unsafeSTToIO -- "I guess that's called \"runST . unsafeCoerce\"" |
| 2025-02-25 11:22:24 +0100 | [exa] | goes aFfInE TeNsOrS |
| 2025-02-25 11:21:49 +0100 | <[exa]> | ski: anyway the main thing I took home from the AD currently is that the array indices need some completely different approach before this works automatically in all cases |
| 2025-02-25 11:21:11 +0100 | <[exa]> | ski: the differential lambda calculus is btw not directly related to the datatype differentiation (as with zippers), right? |
| 2025-02-25 11:19:47 +0100 | alfiee | (~alfiee@user/alfiee) alfiee |
| 2025-02-25 11:18:26 +0100 | misterfish | (~misterfis@h239071.upc-h.chello.nl) (Ping timeout: 252 seconds) |
| 2025-02-25 11:14:08 +0100 | gmg | (~user@user/gehmehgeh) (Quit: Leaving) |
| 2025-02-25 11:14:00 +0100 | <ski> | nice :b |
| 2025-02-25 11:13:41 +0100 | xff0x | (~xff0x@fsb6a9491c.tkyc517.ap.nuro.jp) (Ping timeout: 248 seconds) |
| 2025-02-25 11:12:33 +0100 | <tomsmeding> | ski: I'm doing my PhD in the group that's currently doing research on Accelerate, so in that sense I'm in the right place :p |
| 2025-02-25 11:10:10 +0100 | <ski> | ([exa] : you mentioning updates (and reconstruction) made me wonder) |
| 2025-02-25 11:10:09 +0100 | <ski> | mhm |
| 2025-02-25 11:09:58 +0100 | <Athas> | No, that came out of a different line of research. |
| 2025-02-25 11:09:13 +0100 | <ski> | Obsidian ? |
| 2025-02-25 11:08:52 +0100 | <Athas> | Accelerate itself is also a DPH spinoff. |
| 2025-02-25 11:08:23 +0100 | <ski> | mm, i see |
| 2025-02-25 11:07:54 +0100 | ljdarj | (~Thunderbi@user/ljdarj) ljdarj |
| 2025-02-25 11:07:48 +0100 | <Athas> | ski: the things that didn't work in DPH are dead (mostly the vectorisation transform), but a lot of the other ideas are well and alive, or its successors are. DPH begat Repa, which inspired massiv, which to my knowledge is still good and living. |
| 2025-02-25 11:06:33 +0100 | <Athas> | tomsmeding: yes, those plots are from gradbench (but not fully automated yet). |
| 2025-02-25 11:06:07 +0100 | ljdarj | (~Thunderbi@user/ljdarj) (Ping timeout: 244 seconds) |
| 2025-02-25 11:02:51 +0100 | <[exa]> | ski: as in AD in this case |
| 2025-02-25 11:01:10 +0100 | <ski> | [exa] : differentiated as in AD, or say as in differential lambda calculus ? |
| 2025-02-25 11:00:20 +0100 | <ski> | mainly the parallelism, i suppose, tomsmeding |
| 2025-02-25 10:59:53 +0100 | divya- | divya |
| 2025-02-25 10:58:38 +0100 | divya- | (divya@140.238.251.170) divya |
| 2025-02-25 10:55:43 +0100 | <[exa]> | ski: I was trying to find some such connection before and it doesn't really seem easily. IMO we'd need some very interesting encoding of the array indexes and updates that can be differentiated directly and then reconstructed. Most differentiation formulations I've seen basically deny that. |
| 2025-02-25 10:53:54 +0100 | kuribas | (~user@ip-188-118-57-242.reverse.destiny.be) kuribas |
| 2025-02-25 10:46:42 +0100 | <tomsmeding> | in any case, the fact that it doesn't really exist any more makes this a hard proposition :) |
| 2025-02-25 10:46:27 +0100 | <tomsmeding> | ski: are you thinking of just the parallelism part, or also the distributed computing part of DPH? (IIRC they had distributed execution over multiple machines as a core design goal) |
| 2025-02-25 10:45:59 +0100 | tzh | (~tzh@c-76-115-131-146.hsd1.or.comcast.net) (Quit: zzz) |
| 2025-02-25 10:45:24 +0100 | <tomsmeding> | Athas: do you get those nice plots out of gradbench? |
| 2025-02-25 10:45:18 +0100 | <ski> | Athas : wondering whether there'd be any hope of integrating it with something like the above |
| 2025-02-25 10:43:03 +0100 | <tomsmeding> | Real-world impact is not particularly correlated with publication-worthiness |
| 2025-02-25 10:42:36 +0100 | <tomsmeding> | This is at most a workshop talk! |
| 2025-02-25 10:41:51 +0100 | <tomsmeding> | but then it ends up, in this particular niche of haskell AD libraries, improving the state of the art by >500x. |
| 2025-02-25 10:41:37 +0100 | <tomsmeding> | "this is trivial application of the stuff we already know" |
| 2025-02-25 10:41:25 +0100 | <tomsmeding> | it's also in a way unsatisfying. Writing this ad-dual thing is mostly like "this is easy stuff, the interesting research is in tackling harder problems like actually differentiating higher-order code first-class, or as a source-code transformation, or stuff like that" |
| 2025-02-25 10:40:41 +0100 | <tomsmeding> | well, first I have to revise a journal paper which is due in 2 weeks or so. :P |
| 2025-02-25 10:40:19 +0100 | <tomsmeding> | it's funny how people put those graphs in papers, but then the underlying implementations are actually kind of crap :P |
| 2025-02-25 10:40:11 +0100 | <Athas> | No no, I think this is definitely what you should be workign on. Keep up! Don't get distracted! |
| 2025-02-25 10:39:58 +0100 | <tomsmeding> | hah |
| 2025-02-25 10:39:54 +0100 | <tomsmeding> | anyway this is very much a side-track to what I should actually be working on these weeks. But I wanted to get some ideas from you anyway :) |
| 2025-02-25 10:39:36 +0100 | <Athas> | The PyTorch code in ADBench was also crap, but a colleague of mine was so disgusted that he rewrote it, when we needed it for a paper a few years back. |
| 2025-02-25 10:38:56 +0100 | <Athas> | ski: what about DPH? |
| 2025-02-25 10:38:47 +0100 | <tomsmeding> | also pytorch is like "I'm faster than all you guys but you just need to give me data" |
| 2025-02-25 10:38:13 +0100 | <tomsmeding> | heh okay |