• DeepDream Images thread: LSD is obsolete
    24 replies, posted
[B]What are deepdream images?[/B] [IMG]http://i.imgur.com/LOh5Y58.jpg[/IMG] Deep Learning is a new field within Machine Learning. In the past 4 years researchers have been training neural networks with a [URL="http://343hz.com/content/images/2014/Jan/multilayer_1-1.gif"]very large number of layers.[/URL] Algorithms are learning how to classify images to a much greater accuracy than before: you can give them an image of a cat or a dog and they will be able to tell the difference. Traditionally this has been nearly impossible for computers but easy for humans. Deep Learning algorithms are trained by giving them a huge number of images, and telling them what object is in each image. Once it has seen (e.g.) a hundred types of dog heads 1000 times from a hundred angles, it has been 'trained'. Now you can give it new images and it will spot dog heads within the images, or tell you that there are none at all. It also can say how unsure it is. It was always hard to tell what the algorithms were 'seeing' or 'thinking' when we gave them new images. [URL="http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper-into-neural.html"]So in June 2015 Google Engineers released a method for visualizing what the algorithms saw[/URL].. Towards the end of June 2015 they[URL="https://github.com/google/deepdream"] released their code[/URL], so people could see what the trained neural networks were seeing on any image they wanted. We created this sub to put these images in. It also is fast becoming the place to discuss techniques/methods and try out totally new ideas, such as video: [video=youtube;oyxSerkkP4o]https://www.youtube.com/watch?v=oyxSerkkP4o[/video] [B]How do I make my own?[/B] Read more from this reddit post I totally ripped off: [url]https://www.reddit.com/r/deepdream/comments/3cawxb/what_are_deepdream_images_how_do_i_make_my_own/[/url] [B]Websites that does it for you[/B]: [url]http://psychic-vr-lab.com/deepdream/[/url] [url]http://deepdream.pictures/static/#/[/url] Note: These sites have been getting clogged up lately so they might not be working at times. Chances are you have to use this system on your own PC. Also this thing really likes dawgs.
With that Psychic-VR site does it show the original image while I wait for the result? If not then I got the lamest result.
I've seen lots of these images now and the general theme seems to be "everything is dogs". Why is that?
I keep seeing dogs in most of these images that don't even have dogs. [editline]7th July 2015[/editline] People are also uploading a bunch of furry and mlp porn to these sites. Lots of nsfw stuff being uploaded.
[img]https://storage.googleapis.com/dreamscape-bucket1/output/926b6336-3aee-47d1-b823-a530cf8563f0.jpg[/img] These are great
They released the code? Awesome!
Also how would one set this up to use on a computer? Rather do it personally than do it through a website.
[QUOTE=slayer20;48144402]I keep seeing dogs in most of these images that don't even have dogs. [editline]7th July 2015[/editline] People are also uploading a bunch of furry and mlp porn to these sites. Lots of nsfw stuff being uploaded.[/QUOTE] I'm assuming it's because the neural networks have mainly been trained on dogs and cat, therefore the neural network tries to identify dogs and cats in those pictures. The visualizer then shows where in the image it finds the dogs.
Why is every DeepDream image I see always covered in eyeballs?
[QUOTE=Lunik;48153072]Why is every DeepDream image I see always covered in eyeballs?[/QUOTE] Yeah, they're basically all just eyeballs for me too. It's not very interesting if that's all it can find. :v:
someone should host a server for FP and download all the avatar images as the training set, then place a queue to people to ask for their avatar to be eldritch truthifired.
[QUOTE=CoolCorky;48144350]I've seen lots of these images now and the general theme seems to be "everything is dogs". Why is that?[/QUOTE] GoogLeNet is an image classification network made for a competition that had the sub-challenge of distinguishing between many different dogs. For this reason, several hundred out of the 1000 image [I]categories[/I] were kinds of dogs.
[Img]http://i.imgur.com/PmMCq6t.gif[/img] Went to graffiti park @ Austin. Had someone deep dream it up
Eyeballs are reflective, the algorithm might just be trying to match the specular reflections with those commonly seen on eyeballs.
[QUOTE=Craptasket;48153982][Img]http://i.imgur.com/PmMCq6t.gif[/img] Went to graffiti park @ Austin. Had someone deep dream it up[/QUOTE] why is there a g-g-ghost on top of the wall?
Because I post in the subreddit a record label is trying to commission me to deepdream a music video hahahahahah
[QUOTE=DrTaxi;48158763]Because I post in the subreddit a record label is trying to commission me to deepdream a music video hahahahahah[/QUOTE] Considering they probably don't have people who can easily get it to run, that sounds like a mutually beneficial exchange. If you have the resources for it it's probably a very good idea.
Yeah I'll do it if what they're asking is reasonable. Given that processing takes several seconds per frame at least even on a fast GPU, with simple settings, and potentially several gigabytes of VRAM (depending on resolution), it might very well not be. I'm pretty sure I couldn't process 1080p, neither at home nor EC2. I have a sneaking suspicion that the script isn't as fast as it could be - and it definitely leaks memory, and the convnet implementation Caffe uses is several times slower than the fastest, though that's handwritten assembly, runs only on Maxwell GPUs and would be rather difficult to interface with - but I wouldn't expect miracles either.
I've been trying to figure out for some time to compile this and try this out on my own computer, spec-wise it's pretty good. However, I can't for the life of me figure out how to. Anyone here happen to know how to get it to work?
I didn't ace my machine learning class, but I know this is really useful. It would take me a while to rewrite the code from stratch in something both more robust and more optimized, even then I'm not sure how great the code will be, (NN, parallel programming, and it's new so I don't know WTF some of it is).
[QUOTE=Ithon;48160878]I didn't ace my machine learning class, but I know this is really useful. It would take me a while to rewrite the code from stratch in something both more robust and more optimized, even then I'm not sure how great the code will be, (NN, parallel programming, and it's new so I don't know WTF some of it is).[/QUOTE] The deepdream-specific code is not exactly a lot (or all that performance-relevant), and you probably don't want to implement a convnet all by yourself. What might make sense though is implementing the deepdream parts on a lower level.
[img]https://deepdreamit.s3.amazonaws.com/images/5429/deepdream_c086bac1ded7183ef9cbf722271764ea.jpg[/img] DreamsTime meets DeepDream.
[url]http://deepdream.pictures/imageview?uuid=input/81380dea-dddb-475c-924b-809b0499beea.jpg[/url] [url]http://deepdream.pictures/imageview?uuid=input/62c6fba4-766d-4b80-9c93-cbedd7532ef7.jpg[/url] has anyone deep dreamed the famous cat painting? I wonder if there is going to be just dogs pop up.
War: [IMG]http://i.cubeupload.com/8MinDH.jpeg[/IMG] Rocket launch failure: [IMG]http://i.cubeupload.com/eeyrZ4.jpeg[/IMG] Space: [IMG]http://i.cubeupload.com/yAMZhm.jpeg[/IMG] GTA V: [IMG]http://i.cubeupload.com/tkHOes.jpeg[/IMG] [IMG]http://i.cubeupload.com/ePDWwf.jpeg[/IMG] I'm fairly obsessed with this now.
Sorry, you need to Log In to post a reply to this thread.