Hey Veeky Forums, I'm creating a machine learning model to classify pictures and predict the market based on the ratio of bog/pepe/wojak/other posts.
Send me a picture to test it on. It's still a work in progress. If the pic you send is one of the above it'll get added to my training data.
Pic related, percent chance the model thinks it's in a category.
Attached: Output_screenshot_18.03.2018.png (800x754, 72K)
March 18, 2018 - 12:02
Attached: jo.jpg (3200x1680, 578K)
March 18, 2018 - 12:12
The bog/normal data definitely needs the most work.
Attached: john_oliver_judged.png (800x420, 451K)
March 18, 2018 - 12:17
Attached: 1520795442578.png (670x700, 295K)
March 18, 2018 - 12:18
Attached: guesswho.jpg (259x194, 15K)
March 18, 2018 - 12:18
Gonna add this to both the pepe and wojak datasets.
Deformed wojaks like that are harder I guess.
Attached: together.png (800x835, 381K)
March 18, 2018 - 12:20
The outline of the merchant looks a lot like Pepe, actually.
Attached: merchant.png (800x599, 282K)
March 18, 2018 - 12:22
Bump, give me more data fuckers.
March 18, 2018 - 12:26
>market sentiment analysis for images >no sminem section
don't waste your time breh
March 18, 2018 - 12:26
Give me some smimen and i'll add him in.
March 18, 2018 - 12:28
congrats im gonna make a program that randomly post pink wojaks when the price is going up and green nujaks when the price is going down hahaha faggot
March 18, 2018 - 12:29
Please do. I can grab the signal out of that.
March 18, 2018 - 12:30
Attached: 1514925556528-biz.jpg (500x750, 64K)
March 18, 2018 - 12:31
sure you can kiddo ill murky your signal more than you could ever imagine i graduating top of my class of software engineering btw
March 18, 2018 - 12:33
Attached: got_it.png (800x1200, 943K)
March 18, 2018 - 12:34
Got any good jobs lined up? >Implying GPA matters.
March 18, 2018 - 12:35
Attached: berry-valentine.jpg (282x299, 22K)
March 18, 2018 - 12:36
Attached: 1521388623223.jpg (1080x1349, 85K)
March 18, 2018 - 12:37
>i graduating sounds like its from a top university, user
March 18, 2018 - 12:38
Attached: fuck.png (800x848, 270K)
March 18, 2018 - 12:39
Attached: it_can_tell.png (800x999, 600K)
March 18, 2018 - 12:41
Attached: thinks_its_a_wojak.png (800x771, 352K)
March 18, 2018 - 12:42
you need special brrrrrrrrrrrrrrrrrrrrrap and trap categories for Veeky Forums
Attached: carlos.jpg (1280x720, 73K)
March 18, 2018 - 12:42
Attached: nothing.png (800x450, 332K)
March 18, 2018 - 12:44
Attached: usdt.png (200x200, 15K)
March 18, 2018 - 12:46
Attached: tether_confirmed_pepe.png (800x800, 28K)
March 18, 2018 - 12:47
Now the real question since 50% of shitposts on Veeky Forums are about it.
Attached: linky.jpg (694x800, 71K)
March 18, 2018 - 12:48
also i'm quite certain this wojak will be pepe
Attached: 1521103501572.png (646x700, 166K)
March 18, 2018 - 12:49
Attached: Fibwojack.jpg (1217x1685, 1.12M)
March 18, 2018 - 12:50
>Nothing
Attached: stiny_linky.png (800x922, 384K)
March 18, 2018 - 12:52
Attached: 1474924931266.jpg (761x915, 107K)
March 18, 2018 - 12:52
Attached: oold5ygcwlcy.jpg (540x378, 34K)
March 18, 2018 - 12:53
>bog 0% never ever above $1 confirmed
Attached: 1520578596610.jpg (472x472, 27K)
March 18, 2018 - 12:53
Attached: 1417862633609.png (1000x1000, 51K)
March 18, 2018 - 12:54
can you make a program that counts how many times a coin is mentioned to see which coins are being talked about most and least on Veeky Forums?
March 18, 2018 - 12:54
Can you make a trading bot that takes the latest 100 images on Veeky Forums and decides whether to buy or sell depending on the Pepe/Wojak ratio.
Attached: pard.png (241x241, 60K)
March 18, 2018 - 12:56
You're right, but I was close.
Attached: greenie.png (800x866, 217K)
March 18, 2018 - 12:56
Attached: def_confirmed.png (800x560, 563K)
March 18, 2018 - 12:58
would it be better for the machine to concentrate on the top 100 used images of pepe and wojak? Pareto principle.
March 18, 2018 - 12:59
That's the idea.
Attached: baby_pepe_wojak.png (800x800, 205K)
March 18, 2018 - 12:59
Oops didn’t see other reply.
March 18, 2018 - 13:00
The images often get edited slightly, but I may weigh those images more based on how often they appear. I need a lot more training data.
March 18, 2018 - 13:00
See if it recognizes this.
Attached: 1519487972459.jpg (904x583, 325K)
March 18, 2018 - 13:02
Here's an easier one.
Attached: 1517849062966.jpg (225x224, 9K)
March 18, 2018 - 13:03
Attached: kamikaze_pepe_judged.png (800x515, 462K)
March 18, 2018 - 13:03
Here's a bit harder.
Attached: 1521069357156.jpg (960x720, 501K)
March 18, 2018 - 13:04
After I threw it in azure's computer vision API
{ "tags": [ "plane", "airplane", "flying", "skiing", "air", "small", "smoke", "red", "jet", "blue", "snow", "white", "man", "riding" ], "captions": [ { "text": "a plane flying in the air", "confidence": 0.5739788 } ] }
March 18, 2018 - 13:04
Attached: 1521351708059.jpg (485x443, 40K)
March 18, 2018 - 13:04
Impressive.
Attached: 1514739638927.jpg (900x900, 77K)
March 18, 2018 - 13:05
Attached: wrong8.png (800x796, 194K)
March 18, 2018 - 13:05
Seems like the change in outline throws it off.
Attached: furry_wojak_judged.png (800x600, 629K)
March 18, 2018 - 13:06
Attached: judged_grug.png (800x730, 295K)
March 18, 2018 - 13:07
Strange that it recognized the plane one so accurately.
Maybe it was the coloring.
Attached: 1518919577788.jpg (466x615, 65K)
March 18, 2018 - 13:08
Attached: 100_percent_pepe.png (800x800, 479K)
March 18, 2018 - 13:09
Does it false-positive smug anime faces as wojaks?
Attached: C9ZlmpwWsAEP7ff.jpg (478x380, 27K)
March 18, 2018 - 13:09
The pictures are a response to the market movements, so unless you implicitly conjecture that Veeky Forums and the markets are an effective feedback loop, you're basing predictions off of reactions to past events which makes no sense.
March 18, 2018 - 13:10
Try this.
Attached: 1498261440064.png (280x291, 88K)
March 18, 2018 - 13:10
Yeah, guessing color. Not a hint of green there.
Post more bogs fuckers.
Attached: tritium.png (800x1055, 922K)
March 18, 2018 - 13:10
Almost.
Attached: almost_wojak.png (800x635, 399K)
March 18, 2018 - 13:11
Definetly a wojak
Attached: FOR_SURE.png (800x831, 245K)
March 18, 2018 - 13:13
This, the pictures usually imply some sort of emotion behind it.
I.e. there's a difference between a suicidal pink wojack and a green wojack with arrows flying behind him
March 18, 2018 - 13:13
Yeah. I'm probably gonna add more subcategories.
desu my plan is less to trade with it and more to see if the ratios actually correlate with market movements over time, just as a side project.
March 18, 2018 - 13:14
Does he false-positive as a bog?
Attached: benedict-cumberbatch.jpg (2034x2940, 1.01M)
March 18, 2018 - 13:16
Green must have thrown it off.
Alright, give me a minute. Going to re-train it.
March 18, 2018 - 13:17
Shit, forgot image.I'll do it next.
Attached: strang.png (800x888, 235K)
March 18, 2018 - 13:20
Confirmed descended from ancient bloodlines
Attached: confirmed_cumber.png (800x1156, 1005K)
March 18, 2018 - 13:22
Attached: 1515263652097.jpg (400x400, 27K)
March 18, 2018 - 13:25
Attached: h6nqzAAq_400x400.jpg (400x400, 19K)
March 18, 2018 - 13:26
I think that's it for tonight, thanks for the data!
May return later this week with more refined models.
March 18, 2018 - 13:39
Well, it's pink.
How does your algorithm work, OP? Do you detect the most common color in the picture? Or maybe feed the bot some hashes of common images (which I would find very troublesome)?
Attached: wojak_wilde.png (500x480, 119K)
March 18, 2018 - 13:59
add mutts to algo
Attached: 20f947c0a41e6ed495867441e9d2c16c.jpg (600x600, 38K)
March 18, 2018 - 14:08
Neural network.
Just multiple layers.
The whole point is I don't know how it works. I just feed it the training data.
March 18, 2018 - 14:08
how about dis?
Attached: nolonger.jpg (800x600, 138K)
March 18, 2018 - 14:12
Attached: index.png (229x220, 8K)
March 18, 2018 - 14:16
Attached: 1519275219019.jpg (512x512, 59K)
March 18, 2018 - 14:18
Attached: 1521185698382.jpg (409x395, 56K)
March 18, 2018 - 14:19
Attached: 312fj.png (441x429, 278K)
March 18, 2018 - 14:19
Attached: IMG_6662.png (960x1280, 1.26M)
March 18, 2018 - 14:21
I already did this, it basically parses OP images in the catalog and returns a coefficient which evaluates overall market feeling
As a friendly pointer, bog posts provide no meaningful insight or data
March 18, 2018 - 14:25
You are what Clif High aspires to be.
March 18, 2018 - 14:32
Attached: 1521273698344.jpg (893x1024, 130K)
March 18, 2018 - 14:35
Attached: 1519535910600.jpg (720x720, 47K)
March 18, 2018 - 14:35
Nice; what features/images did you end up parsing out as important?
If you want to share, ofc.
March 18, 2018 - 14:38
How large is your current data set and how long did you spend preparing it?
March 18, 2018 - 15:39
make the bot run every 30 minutes or so and have it post the results to a private telegram channel that us Veeky Forumslets can see. this is an interesting project desu
Attached: IMG_2141.jpg (3200x3200, 888K)
March 18, 2018 - 15:44
~100 images each. I need more. Not very long.
March 18, 2018 - 16:07
Share the link of the GitHub repository OP. Once you are done
March 18, 2018 - 16:10
Try this
Attached: 37d.png (1037x1037, 1.09M)
March 18, 2018 - 18:15