« first day (2757 days earlier)      last day (2385 days later) » 

00:00
I'm thinking of doing that I can't break 88% val acc..
how did you get 100% test accuracy and only 88% val??
All of my models that had 88% validation accuracy had like 80% test accuracy, they were garbage
@quartata I assume some of the validation test cases were nonsense titles
I had 1000 validation cases roughly
oh, that's a huge validation set
yeah, I don't see any point in decreasing it though since I'm overfitting quite a bit after 2 epochs
...2 epochs???
what on earth is going on here
00:03
Epoch 1/5
 - 89s - loss: 0.5730 - acc: 0.7395 - val_loss: 0.3815 - val_acc: 0.8539
Epoch 2/5
 - 88s - loss: 0.3193 - acc: 0.8834 - val_loss: 0.3332 - val_acc: 0.8659
Epoch 3/5
 - 88s - loss: 0.2395 - acc: 0.9222 - val_loss: 0.3226 - val_acc: 0.8749
wait a second I have goofed
what happened?
Anonymous
@Downgoat Could you add full code for your solution for testing/validating? When you undelete it, that is
@Fatalize so like Charcoal :P
@Mego yes
Anonymous
Thank
00:09
@quartata I don't know how to python is what happened
@quartata what is wrong with two epochs
It's just not very long is all
to start getting significant overfitting
Although yours doesn't look that bad, the validation loss was still going down
@Downgoat Your model outputs True for all examples, so it's 50% on the test set
Classic overfitting. Sorry.
Er
Oops sorry I forgot that yours outputs two
00:25
@EsolangingFruit now time to tackle the second top unanswered question
@quartata can’t be right, I did the percentage wrong I am aware which is why I deleted but it’s around 80%
Yeah that was my bad
actually
Are you sure?
I'm still getting 50%:
@Downgoat link pls
model = keras.models.load_model("downgoat.h5")

vocab = __import__("pickle").load(open("/Users/quartata/OnionNotOnion/words.pkl","rb"))

tp = 0
fp = 0.0

with open("list", "r") as fh:
    for i, headline in enumerate(fh):
        arr = numpy.zeros((1,59))

        for j, token in enumerate(STRIP_PUNCTUATION.sub("", headline.lower()).split()):
            try:
                arr[0,j] = vocab[token]
            except KeyError:
                arr[0,j] = 0

        result = model.predict(arr)[0]
pruned but
The outputs are usually along the lines of [ 0.68659937 0.31340063]
And inputs look good to me:
[[  2499.   7968.  13245.   1122.   7048.   2878.  12961.   5140.  12850.
   17277.   6911.   5363.   1122.   5612.  14597.      0.      0.      0.
       0.      0.      0.      0.      0.      0.      0.      0.      0.
       0.      0.      0.      0.      0.      0.      0.      0.      0.
       0.      0.      0.      0.      0.      0.      0.      0.      0.
       0.      0.      0.      0.      0.      0.      0.      0.      0.
       0.      0.      0.      0.      0.]]
25 25 0.5
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
embedding_1 (Embedding)      (None, 59, 48)            1051152
_________________________________________________________________
spatial_dropout1d_1 (Spatial (None, 59, 48)            0
_________________________________________________________________
lstm_1 (LSTM)                (None, 36)                12240
does that look right?
@quartata IIRC pad_sequences pads 0s from the beginning so I'd so that (idk if that makes a difference)
00:35
Oh, so it takes them in reverse
?
(A lot of LSTMs do that which is why I ask)
Oh nevermind I see what you mean
yeah it's like:
idk how to copy paste from notebook :(
OK I got 78%
39, 11
Makes sense, it's probably forgetting everything by the time it gets through the pads
Mine doesn't go like that so
@quartata pads?
the zeros
wait a second h*ck i did a stupid
00:40
hm?
Wondering what this 273-word headline is that's in my training data apparently
@quartata from reddit? I don't think they are any 273 word headlines..
...oh dear
um
I accidentally used character length
so these are way too big
OK this explains a couple things
Well, we're at a fresh start here
New data, tweaked the model
time to redeem myself and look like I actually know how to do my job
@quartata how long does it take you to train your model
I have a couple different architectures
this one has two phases
the first phase, maybe 15-20 minutes
the second phase, rest of the day although usually only a couple epochs are good
I let it run as long as possible because I have set up only to save the best anyways
so it doesn't hurt to do a lot of epochs
I didn't really bother to tweak my minibatch size mind you
probably could make it go faster
meh
zzzz pushshift is so slow
@Downgoat Interesting, I'm wondering now where your 59-word headline came from because the biggest I have is 45
And I'm pretty sure I've got it right now :P
00:55
o_o
@quartata O__o really I'm able to get like almost all posts in 10s of seconds
You can only get 1000 from the reddit API
I'm having to use this other thing to get more
01:08
@quartata yes I am using pushshift if that's what you're using
Yeah
maybe that's why it's taking so long because we're both slamming it
use the elasticsearch backend if you're not already
Although I'm only asking for the IDs from it
using PRAW to get the actual data
wait why
to be honest I don't know
just what I wrote
it's probably a bit nicer to pushshift
@Mego wait, 20% or 20 percentage points
(both seem too generous if you ask me but)
All righty, 2000 onion 5000 not the onion
Starting from scratch
Anonymous
01:40
@quartata 20 percentage points. So if you get an 80 on the public cases, you have to get at least a 60 on the hidden cases.
all right
01:53
Man, this model has 91% validation accuracy but only 80% test accuracy. What gives.
I blame Mego's test set
I probably can't fetch more data so I'm kinda pondering what to do, I doubt going deeper will help
I only ran 6 epochs of finetuning but it wasn't going anywhere better
Oh, I had a small padding bug, it's actually 82%. Still.
no better than the solutions that search for "Dad" (???)
02:15
Ugh, too big for Github.
Hm hm hm
02:39
@Mego You around to give me my scores on the hidden set?
I reluctantly posted mine even though I'm not satisfied yet
Anonymous
@quartata Yep, give me a minute
Anonymous
02:55
Oh jeez keras is not small
Sorry :/
My model is 250 MB also
Big embedding
W W
W W
03:38
@JoKing You might want to update your bio. You've done some pretty awesome things recently. Things I certainly would be proud of.
@DJMcMayhem Can you unfreeze the esolangs room?
Anonymous
@WW Poof
W W
W W
thansk
03:58
@Mego did you get Keras working?
Anonymous
Not yet
Anonymous
Dependency hell
@quartata I ran into GitHub's filesize limit the other day when I tried to upload the Clang binary, which, unstripped, ended up being 2 gigabytes.
How can it be so huge
@quartata No clue, but I stripped it and it become 200 megabytes
Which, after DEFLATEing it, got it under the limit
...IDK why I put deflate in caps there
LZMA Deflate
04:14
@quartata O_O WTF
how many trainable params?
Don't remember off the top of my head
probably a lot
04:56
why the hell do you people think the cinnamon stick post is funy
the only way to appease you baboons is slapstick self-harm
abhorent
@Mego to test the model just clone the github repo and then put all the code from my post in order in a python file and then get_type will return from which category that is. I can write a wrapper script w/ stdin if that'd help
Your code doesn't work out of the box
len(words) isn't defined but if it was it would include the UNK symbol which seems to make it off by one
05:12
@quartata wait what
H*CK
ok well I made a better model so I'll update fix with that
ok fixed with new model
Added a test script to README for info
FYI: Python 3.7 doesn't work with Theano backend
Turns out another dense+layers plus dropouts basically removed all overfitting
 
3 hours later…
08:44
0
Q: Generate N digits of pi with python

MikaellangeloI have tried to write a program that generates digits of pi using the CHUDNOVSKY ALGORITHM to calculate digits of PI (https://en.wikipedia.org/wiki/Chudnovsky_algorithm). In my code I read in a text file with the first 100000 digits to compare it to. My problem is that is only seems to be able to...

 
1 hour later…
10:04
help.
can i do scanf("foo") (i.e. no format specifier) and check for success/failure
10:21
I don't think so?
10:33
@quartata marky markov
@dzaima because "downwards extending" is actually creating a new cursor (I think?)
@ASCII-only yes it is, but that doesn't excuse the annoyingness
@dzaima it's not annoying...
without alt is moving the endpoint
with a single selection and no alt, it'd shift the endpoint down by one
which usually doesn't make sense for multiple cursors
@ASCII-only it is at least way more annoying that Notepad++'s way
@dzaima and how is notepad++ any different
and you can't shrink the box vertically which is another horrible thing
10:44
oh
solution: vim visual mode
with onivim
@ASCII-only shift+alt+right/left is the same as shift+right/left but for the box mode
@ASCII-only i don't want to waste time learning vim
@dzaima onivim is graphical
basically sublime but 100000x better and open source (i.e. not nagware)
also npp allows selecting the bounaries outside of the text and padding with spaces where necessary which is very nice
@ASCII-only then why is vim in its name
@dzaima ... because it's vim based?
like gvim but less focused on the vim part
11:02
@ASCII-only finally got it installed and yep it's vim.
@ASCII-only I thought basically sublime would at least include the regular ctrl+ things and that mouse dragging would select text
@dzaima wait, it doesn't?
@ASCII-only on it seems it selects whole words (was testing on a single like of alphabet spam) but the cursor jiggles back and forth from the mouse position and the words start/end
uhh how can I even paste something not copied from within it
@dzaima i -> ctrl+shift+v probably
also set mouse=a in .vimrc to enable mouse. although i should really try onivim shouldn't i :thinking:
ctrl+v by some reason started working, though sometimes only after a mouse click, which, sometimes, enters the text <LeftMouse>... I think i'll just give up on vim, that seems simpler that understanding its mysteries.
11:20
@dzaima :|
@dzaima just set mouse=a
@ASCII-only okay idk what that has changed but still I'd have to learn way too much to be able to just do basic things in oni for me to bother
ಠ_ಠ oni→preferences→edit config→<i> gives "cannot make changes"
@dzaima wait what
10/10
oh that's a reference. I have to write typescript to actually modify the config.....
there's no 1 char/keystroke thing for moving a line up/down by default? not only do I have to relearn all the basic actions but I have to write my own shortcuts too..
11:45
0
Q: Codegolf Rainbow : Sorting Colors with Reflection

Kevin CruijssenIntroduction: After I posted two rainbow-related challenges: Codegolf Rainbow : Fun with Integer-Arrays 1 and Codegolf Rainbow : Draw in Black-and-White 2, the following comment was made by @ChrisM in the ASCII (Draw in Black-and-White) challenge: Maybe you know this and it's by design (I kn...

11:58
:|
Can you repro this?
readline doesn't work with -O in C
thanks a lot, readline
@Mr.Xcoder suddenly, yes; a few minutes ago, no
Same here. IMO they are truly horrendous
the thing is, though, that it should probably be posted over mother meta, since I don't think it's a technical issue
12:02
But I cannot repro it on other SE sites
@Mr.Xcoder no. There is a css text-decoration: none; for me, which is supposed to remove that
I know it is supposed to, but it does not work for me
Clear caches maybe?
9
Q: Underline appearing for hyperlinks

ArulkumarAll the hyperlinks in the Post and Comments in the site are displaying with the underline. Sample link. Previously the hyperlinks are displaying without underline. It can be reproduced in recent Google Chrome and Mozilla Firefox browsers. Is the expected behavior or bug? Screenshot for referen...

Uh...
@Mr.Xcoder I did that, no effect. Though now there are underlines on the meta :|
There is a text-decoration: underline; for me :|
12:15
:| fscanf is segfaulting
C is an insecure function
13:02
help how to svn
13:16
0
Q: Sort by what the digit pairs describe

sundarGiven a positive integer, we can form a new number that's described by its digits taken pairwise (with a leading 0 added for numbers with odd number of digits). For eg.: 1234 can be read as one 2, three 4s - so, the output for 1234 is 2444. 643 has an odd number of digits, so a leading zero ...

@Mr.Xcoder I want to disallow list of lists of digits as input for this question, I feel that makes it a bit less interesting.
But there's already a Python answer with string input, and that brings the "is a string a list or not" mess into it
Shall I just ban string input too and mention that in a comment on that answer ?
How can lists of digits make a challenge less interesting though?
I'm thinking of mainstream languages where integer -> list of digits is not so straightforward
So integer-only input might lead to some clever solutions
If you disallow lists of digits, disallow strings too. Otherwise, allow both. IMO you should go with the latter, but if you really want to ban them, ban both
@Mr.Xcoder I don't think outputting lists of digits is going to shorten your code, you have to sort at some point
13:29
@EriktheOutgolfer Yes it would if input as a list of digits were allowed too.
@Mr.Xcoder I think I'm going with banning them both, I'll update the question and comment on the Python answer
As you wish
ಠ_ಠ You allowed that for output but not for input
@Mr.Xcoder :P Yeah, I actually didn't originally see that you'd asked about both. Using that for output (when that's not the input format) feels like allowing just a possible minor hack, not something that takes away from the challenge.
13:45
I have an idea for a Jelly answer now...
@sundar the "pad with a zero for numbers of odd lengths" is really annoying
Well if there was no sorting and the pad wasn't there I'd hammer it as a dupe of the run length decoding.
@Mr.Xcoder shakes fist damn loophole
That's basically the whole challenge tbh
13:48
@Fatalize The pad comes directly from the OEIS sequence
In fact the whole question is just an OEIS page that I randomly came across and searched for on PPCG
I left it on the Sandbox for over a week because I was sure it had to be a dupe!
And it doesn't because of that twist
14:18
lol at least do it right dude
not that I'll be watching your endless video anyway
14:46
2
Q: Compute the minimum \$a(n)>a(n-1)\$ such that \$a(1)+a(2)+\dots+a(n)\$ is prime (OEIS A051935)

Mr. XcoderBackground Consider the following sequence (A051935 in OEIS): Start with the term \$2\$. Find the lowest integer \$n\$ greater than \$2\$ such that \$2+n\$ is prime. Find the lowest integer \$n'\$ greater than \$n\$ such that \$2 + n + n'\$ is prime etc. A more formal definition: $$a_n=\beg...

15:31
0
Q: Number Spiral Problem

Agile_EagleA number spiral is an infinite grid whose upper-left square has number 1. Here are the first five layers of the spiral: Your task is to find out the number in row y and column x. Example: Input: 2 3 Out : 8 Input: 1 1 Out : 1 Input: 4 2 Out : 15 Note: Any programming language is a...

16:24
@WW in ^ in what way do you see the inputs being 0-indexed?
W W
W W
Yeah they are 1 indexed, the zero and the 1 should be swapped in my comment
There I deleted and readded the comment.
think we accidentally killed Mego trying to set up tensorflow
this means I win by default
16:46
-1
Q: clustering arrangements from a given number of records

ss321cI am working on following problem. Any help in term of algorithm as how can this be achieved will be great.The hint I have is such kind of clustering arrangements are represented by bell numbers this is from where the problem originated in my opinion. Write a program to find all the possible clu...

that was me actually
Anonymous
@quartata The reports of my death have been greatly exaggerated
Anonymous
I've actually been at work, which is worse than death :P
@quartata Yeah I'm convinced
wearing his hollowed out carcass like a complete dumbass while trying to buy lucumas
@Mego hello me
17:09
What's a lucuma?
abdu has them 75% off although I'm starting to think the health department isn't going to appreciate me hanging around food dripping blood
so im going to go later once it dries out a bit
Have you considered not wearing a hollowed out carcass
no
why
So that you don't drip blood everywhere
17:17
eh
17:28
Oh, chat... I leave for a couple of hours and now you've spiraled into hollowed out carcasses and dripping blood
5
I like it.
17:56
@Adám I do find that to be useful
18:13
@J.Sallé if it helps i took it off
it was getting too hot, it was like 4K outside
@quartata wait what
@Mego for testing python answers you can use google colab— has tensorflow pre-installed
Yo @Downgoat, my userscript doesn't load the goats when I activate goat mode. Is it a thing that's happening or did I mess something up?
@J.Sallé uh I think goat alignment might be messed up but I will work on it when I'm back home next week
@Downgoat oh okay, I thought I might've messed something up somehow. I'm a big dumdum when it comes to JS
18:32
0
Q: Arrange vowels in ascending order in the places of vowels from input string

Kodanda Rama Durgarao PoluriGiven a string, we need to output the modified string with the following rules. Arrange vowels in ascending order in the places of vowels Eg: Input: stackexchange Output: stackaxchenge Explanation: we have vowels in indeces 2 - a, 5 - e, 9 - a, 12 - e we arrange them in indeces of vowe...

19:01
Look at this awesome keyboard:
Can you get anything like this for Python (at least an installable one)?
0
A: Sandbox for Proposed Challenges

Conor O'BrienChain Classification answer-chaining related Objective: Write a program (whose index in the chain of answers is \$n\$) which, when given any program with index \$i\$ (\$ 1 \le i \le n\$), outputs \$i\$. (The program may do anything else given any other string input, including but not limited ...

@BetaDecay how many keys though?
im so online right now i need someone to dump a big bucket of gatorade on me to cool down
Sports references? I think you're in the wrong place
what are sports
19:09
@quartata It's when you run around a lot and get tired
It's pretty lame
oh ok that's the stuff jon bois keeps talking about
got it
/me doesn't know who jon bois is
@jon_bois, brooklyn/louisville
well it's jon. waggener high, class of 2001. creative director, sb nation labs
39.3k tweets, 109k followers, following 419 users
Jon-Boy is the guy from the Waltons isn't her?
you all have had your culture stolen from you....
no taste
19:15
@quartata I thought you meant jontron for a moment
@Riker absolutely the hell not
yea I was like "you still watch him?"
i'm actually kinda dissapointed he turned out that way, his videos were only like 45% bad
What happened?
he's racist.
Fame and fortune happened. Also, racist.
19:17
well
Check transcript I think quartata mentioned it before
Lol "genuine Victorian" sounds like they'd be a hipster
19:58
@Riker since you're obviously the person who asked about Nethack Legacy on my curiouscat: is this literally just 3.4 with a crt television shader
 
1 hour later…
21:07
@quartata No, it's 3.6.1, and it also comes with an options menu
Other than that, yes
ais has put out a statment asking everyone to not buy it
@quartata afaict yes basically
and I don't remember doing that but ok
Can someone explain to me how nymphs in nethack weren't a horrible idea
@Riker never heard of that series, I probably need to know it to get what's going on?
@Pavel that's my second geno
L then n
also go play dnh engagment ring OP
@flawr not really, the point is that the show writer made a nice long response that culminated in a meme
@Riker I've tried a couple times and had no clue what was going on
21:15
I can walk you through if you want to play rn
can't rn
otherwise it's pretty simple just make sure you ask somebody if you find a wish or a special level
@quartata tweet idea: really stupid rules on when to seed torrents
@Riker sorry, too complicated for me to understand at the current time of the day and the current blood alcohol concentration :)
21:48
0
A: Sandbox for Proposed Challenges

flawrP Pr Pre Pref Prefi Prefix Prefixe Prefixes code-golflistarray Given some finite list, return a list of all its prefixes, including an empty list, in ascending order of their length. (Basically implementing the Haskell function inits.) Details The input list contains numbers (or another ty...

 
1 hour later…
22:52
@Pavel oh, it was you
@quartata i just want you to know I'm giving your twitter free advertising
my town of salem name is now "quartata anatre"
braver than the troops
almost just froze chrome copy-pasting a 3 MB data:image/png URL
it was worth it
23:28
@Riker never watched the series
23:58
I will never understand why APT thinks that the best solutions are the ones that involve uninstalling 20 packages, and the one that entails merely updating 2 only comes 8-9 solutions down the line

« first day (2757 days earlier)      last day (2385 days later) »