|
You. You worked for six months on a structured prediction method. Your message-passing algorithm implicitly maximized approximate entropy over the marginal polytope. It theoretically elegant and has proven convergence guarantees! Learning is a simple matter of gradient ascent on that noisy objective which is exponentially complex to evaluate (lol). It was a nightmare to implement, and it didn't work as well as logistic regression, because gently caress you. Your Model, unfortunately, is a Piece of poo poo. Welcome to the club of actually doing machine learning, where the only thing that's always true is that approximately interesting => probably incorrect. This thread is for people who are currently finding the kitty in the picture or want to learn how. Also, for wildly shitposting when you're trying to figure out what the gently caress is happening in your symbolic computation graph because print statements get optimized out unless that print has a downstream op so you need to set something equal to the result of the print because why on earth would anything ever just work and make sense? Why would you want that at all? Let the tensors flow over you.
|
#
?
May 26, 2017 18:33
|
|
|
|
| # ? May 9, 2024 06:23 |
|
|
|
|
#
?
May 26, 2017 18:35
|
|
|
my model predicts that the best course of action in to "gan thread, bas op" so not quite there but pretty good
|
|
#
?
May 26, 2017 18:38
|
|
|
Deep Dish Fuckfest posted:my model predicts that the best course of action in to "gan thread, bas op" idgi but here's more weird heads  oh wait 
Khym fucked around with this message at 19:12 on May 26, 2017 |
|
#
?
May 26, 2017 18:41
|
|
|
|
|
#
?
May 26, 2017 18:49
|
|
|
https://twitter.com/boredyannlecun/status/862150316724072449
|
|
#
?
May 26, 2017 18:51
|
|
|
|
|
#
?
May 26, 2017 19:14
|
|
|
i like lode runner
|
|
#
?
May 26, 2017 19:42
|
|
|
my model is very bad
|
|
#
?
May 26, 2017 19:46
|
|
|
|
|
#
?
May 26, 2017 19:51
|
|
|
reposting http://oneweirdkerneltrick.com/
|
|
#
?
May 26, 2017 19:51
|
|
|
amazing
|
|
#
?
May 26, 2017 19:54
|
|
|
seriously though where are those face gifs from and is there a paper or something about them?
|
|
#
?
May 26, 2017 19:55
|
|
|
A machine vision system was developed to identify corn kernel breakage based on kernel shape profile for automated corn quality inspection. https://elibrary.asabe.org/abstract.asp?aid=28547&t=2&redir=&redirType=
|
|
#
?
May 26, 2017 19:57
|
|
|
Khym posted:amazing i fuckin love every one of their papers
|
|
#
?
May 26, 2017 19:59
|
|
|
Deep Dish Fuckfest posted:seriously though where are those face gifs from and is there a paper or something about them? it popped up on my twitters today https://twitter.com/prostheticknowl/status/868114244260286464
|
|
#
?
May 26, 2017 19:59
|
|
|
I've never figured out why kernel is so overloaded in ML (and CS in general) wtf does a cnn filter (kernel) have to do with a similarity matrix (kernel)? Why is it the same thing as the thing that talks to hardware for me?
|
|
#
?
May 26, 2017 20:01
|
|
|
in hindsight that OP was a little manic but I think it accurately captures what it feels like to write and debug ML code
|
|
#
?
May 26, 2017 20:02
|
|
|
why are they using that guy from film photography project?
|
|
#
?
May 26, 2017 20:08
|
|
|
|
|
#
?
May 26, 2017 20:12
|
|
|
loving hell https://twitter.com/quasimondo/status/868003676534837248
|
|
#
?
May 26, 2017 20:16
|
|
|
R has several COBOL libraries and dependencies depending o. your modeling, which is especially hilarious when they break
|
|
#
?
May 26, 2017 21:28
|
|
|
Share Bear posted:R has several COBOL libraries and dependencies depending o. your modeling, which is especially hilarious when they break One of the nice things about neural nets is that high-quality engineering organizations have written new code Old stuff eventually calls some fortran library that no one has understood for decades and it's spooky as gently caress
|
|
#
?
May 26, 2017 21:42
|
|
|
chome, but in the style of lomarf
|
|
#
?
May 27, 2017 00:57
|
|
|
Khym posted:Your Model, unfortunately, is a Piece of poo poo. i started building gundam models for a kid at the hospice and i really loving hate it, how do people have patience to sit there and sand and polish and mask and paint poo poo for like 200 hours for fun like, what a waste of time
|
|
#
?
May 27, 2017 02:12
|
|
|
I installed tensor flow and went through the tutorial then I realized that I don't really remember any math and have no idea how to do something useful with this
|
|
#
?
May 27, 2017 02:35
|
|
|
|
| # ? May 9, 2024 06:23 |
|
|
basically you throw some network together using whatever units and architecture your guts tell you would work best for your problem and then you throw the same data at it until it looks like it works but not for too long either or it'll gently caress up when you show it new data if it doesn't work you should probably just throw more new data at it until it does or maybe gently caress with some hyperparameters there might be some vector calculus and linear algebra poo poo in there or something idk who cares
|
|
#
?
May 27, 2017 02:47
|
|