I'll dedicate this rambling thread of words to the linguist Ray Jackendoff who studies semantics + computational theory of mind. This paragraph from his monograph debating the merits of studying the details individual words vs patters across many words is one of my favourites.
Here's a link to my colab if you'd like to give it a go yourself. This codebase builds off of previous work from many people including @advadnoun @RiversHaveWings @NerdyRodent as well as ClipDraw from @kvfrans @crosslabstokyo @err_more and @okw. colab.research.google.com/gi…

Aug 17, 2021 · 12:49 PM UTC

And my next step with these is finding ways of limiting the color palette. If you enjoy this work and would like to contribute (and are comfortable with pytorch) feel free to offer up coding suggestions in this github issues thread. github.com/dribnet/clipit/is…
Replying to @dribnet
I tried it and got this error, not sure what to do.
IIUC - this is what happens if you get an K80 GPU - best bet is to try loading the page again and hope you get an GPU upgrade.
yes - if you edit the settings cell in there you will see "scale=2.5". this can be changed to "scale=4.0" etc. but this uses more memory, so it might crash the system if scale gets too big (memory workaround: can drop quality to "normal" or even "draft")
Did you run the setup cell twice? When In doubt, “Restart and Run All” fixes most things. 🙃
Absolutely minor detail from a layman, but adding a !nvidia-smi -L cell at the top of the notebook helps a lot those of us who don't have Colab Pro and may not get a a T4 or above every run
thanks! I'd like to keep the cells to a minimum but I would certainly be up for adding a check inside the setup that aborts with an alert if the nvidia-smi report back with a wimpy GPU if anyone has a handy code block for this.
Phenomenal! Works especially well with urban/building prompts, probably due to the preponderance of straight lines.