Oh I wish I could code
Is there any VM I can grab with all the dependencies installed?
never mind, will try to install everything on OS X - just hope it won’t mess with my development environment…
Please let me know how it works out pichenettes! I would pay money to have some of my images warped by this software
I have no idea what VM is, or anything about coding, wish I did!
Look at these beauties
Installed Anaconda, brew and all the python dependencies, CUDA 7 still installing…
I wonder if something similar would work with sounds instead of images. ‘This is a snare drum’, ‘this is a trumpet’, ‘this is a Moog’.
Cool, I hope you guys can get it working, I would love to see the results!
@ByteFrenzy, I was actually thinking how this software could be adapted to sound? Would be interesting to hear what computers dream as well as what they see
Yes, I stumbled across this in my work, and I wad immediately starting to think about the possibility of doing audio with these techniques.
It’s not possible to directly reuse this software for sounds.
Behind the few lines of code released in the post above, there is:
- years of research in good network architectures for image recognition.
- months of image database annotation / curation.
- months of tweaking model training strategies (and weeks of computation on large clusters for model training itself).
So we’d need to replicate all these steps first for audio, before reaping the benefits of running networks “backwards” for dreamy audio generation. The use of deep learning for music analysis or speech recognition rarely goes beyond raw acoustic feature extraction (we could make a network “dream” of phonemes or basic waveforms, not evolving sounds and melodies) - most current implementation of speech recognition systems are still using classic FST methods. There are no ready to download nets/models (contrary to what you can find for images)
As someone who does 3D modeling on occasion, how useful is this code for making seamless textures?
Very trippy picture results. Also, you could use the resulting picture in a spectral synthesizer that basically uses a set of rules to interpret a picture to an additive sounding oscillator.
> how useful is this code for making seamless textures?
It will not rearrange the overall geometry of the picture, so I’d say useless.
Still stuck with multiple python versions hell Sad thing is that in 5 or 6 hrs someone will have created a web app running on AWS or the like to do this…
Thanks for the fast response.
I figured as much, but when you look at the generated results on the surface, it seems like the algorithms are adding geometry. IE: A Venus Fly Trap becoming a bird.
No, they cannot do what you suggest. It’s not about piecing image processing algorithms together (and that’s the beauty of the method indeed, since it does everything from scratch).
Let’s take an audio analogy… Imagine you have a magic Behringer mixer with a sophisticated EQ/FX unit and a magic vu-meter that tells you how much the signal it receives sounds like a clarinet. You feed arbitrary audio into this mixer, and mess with the EQ/FX to make the “clarinet-meter” go as high as possible. Now imagine you have meters like this for different categories of sound (a “voice-meter”, a “thunderstorm-meter”, a “fart-meter”…). You could try messing with the EQ/FX to push the reading of one meter as high as possible, and the others as low as possible.
That’s what this code does - it enhances pixel par pixel bits of the image so that the “meters” recognizing various shapes, patterns and ultimately object types read as high as possible. So it doesn’t invent geometry or distort the image - actually, it does the least amount of work as possible to tweak the image so that recognizable shapes emerge from it.
Wasted 2 hours installing this… Now I’m going to waste 20 hours rendering animations
OMG animals start appearing
Woho! Did you need to add this animals as pictures into the network before or is it automatically generating animals out of nothing?
You can download pretrained networks for the image recognition challenges used in academia - which cover many categories (animals, objects, vehicles, scenes from everyday life).
The technique DOES NOT make animals out of nothing - it regurgitates what it has been exposed to during training. This is actually a very good criterion for evaluating the intelligence of any recognition system - run it “backwards” (for a probabilistic model: sample from it) and see what it spits - this is the “essence” of what it has learnt.
That is freaking amazing I just download some software you linked to pichenettes, I hope I can get it working