06:59:42

woke up to let the cat out, she's scared of the front porch? oh there's a guy standing at the screen door. he tells me he can fix my van in the backyard. How does he know it's broken it's not visibly broken? he says someone called him a thief but these bags of trash are his

🗨️ 2 ♺ 0 🤍 12


07:00:52

I say okay man and eventually he leaves. I'm naked So this is a good development. however now I'm wondering what he knows about my van, because it is broken and it developed those broken parts while sitting in that backyard 😡

🗨️ 2 ♺ 0 🤍 13


07:48:36

the world is not going to end

🗨️ 11 ♺ 1 🤍 32


09:15:51

i am the neurotype. everyone else is neurodiverging from me

🗨️ 4 ♺ 7 🤍 62


10:41:59

put my name on a list at your own risk

🗨️ 2 ♺ 0 🤍 6


10:56:10

my meme search project is stuck. thinking out loud here about some of the problems

🗨️ 1 ♺ 2 🤍 22


10:57:56

the last working build of memery was 0.0.8 on pip. after that I started trying to fix the dependencies and instead broke them. so 0.0.9 and 0.0.10 are effed up and i don't know how to remove them from pip

🗨️ 2 ♺ 0 🤍 13


11:00:17

the dependencies were a problem partly because PyTorch and CUDA versioning is always annoying, partly because of the literate programming framework i decided to use, and partly because of the way packaging is handled for the CLIP model. plus i've never been good at virtualenvs D:

🗨️ 1 ♺ 0 🤍 10


11:02:50

pip won't let a package install random shit from Github, which is probably sane behavior. but OA doesn't package CLIP for pip, they recommend you install from Github. i was using a third-party conversion module called clip-by-openai, but it unhelpfully pins PyTorch version

🗨️ 2 ♺ 0 🤍 10


11:09:56

vibe camp

🗨️ 1 ♺ 0 🤍 37


11:14:32

so I started trying to convert to huggingface transformers. only to find that they use a completely different data model for batching? great..

🗨️ 1 ♺ 0 🤍 9


11:16:15

then i tried replacing all my backend code with a Flow from https://t.co/hm8g3hkbzi. this actually worked okay, but to get CLIP encoders and the indexing functions installed I need at least three Docker containers and like 5GB of hard drive space. This won't work for a local app

🗨️ 1 ♺ 0 🤍 9


11:17:45

okay, maybe I can host the model and indexing stuff on a VPS somewhere and make a frontend app to talk to the API. that would allow me to use any computer or phone to save and search the same meme database. tried prototyping this locally -- CORS problems! 🤬

🗨️ 1 ♺ 0 🤍 10


11:19:06

i think i can run an API on one localhost port and the frontend at a different localhost port and have them talk to each other. But since I'm using remote databse now, I need to pipe all the images over the wire? this sucks especially when they're already in a local folder

🗨️ 1 ♺ 0 🤍 8


11:19:47

definitely can't access those local files to display in a local website either! the browser is sandboxed from the file system. again, this is reasonable, but annoying

🗨️ 1 ♺ 0 🤍 7


11:21:26

i can see image blobs that Jina is porting around, but for the life of me I can't figure out how to translate them into something the browser can display! maybe it doesn't help that this is my first clojurescript project. but would native JS/React solve this any better?

🗨️ 1 ♺ 0 🤍 7


11:22:42

now I'm reconsidering the whole 'cloud-native' model. i really like building local image search. and I can think of use cases, like big image datasets, where it would be useful to search and manipulate data locally but wouldn't be economical to upload and do remotely

🗨️ 1 ♺ 0 🤍 5


11:26:07

it's expensive to use CLIP! it takes either a powerful GPU or a lot of time or sometimes both. In my original buildout I use torch dataloaders to do batching functionality but it's still the heaviest step in the pipeline for sure. indexing and storing embeddings helps, of course

🗨️ 1 ♺ 0 🤍 5


11:28:08

but it really seems like a piece of metadata that should be available to the operating system, file picker, etc. once it's been computed at all. what if memery were integrated into GNOME Overview or Spotlight

🗨️ 1 ♺ 0 🤍 5


11:30:45

on the other hand, i was clearly thinking differently about it a month ago https://t.co/EUtpV7Npgt

🗨️ 1 ♺ 0 🤍 7


11:37:54

he made it sound so easy https://t.co/2bG6VBO7G0

🗨️ 0 ♺ 0 🤍 10


12:06:39

Wake up and rage

🗨️ 0 ♺ 0 🤍 5


12:09:10

is there ANY reasonable way to package a PyTorch model for local machines? or is it just up to the user and their OS

🗨️ 4 ♺ 0 🤍 16


12:53:16

can't believe that the best solution for packaging software in 2021 is "include an entire operating system and virtual machine"

🗨️ 7 ♺ 1 🤍 46


13:20:29

What if I told you going to write a CRUD app with jQuery

🗨️ 11 ♺ 0 🤍 31


15:55:21

i have been reading about htmx for two hours

🗨️ 1 ♺ 0 🤍 4


16:12:41

I'll be like I know a spot and then take you here

🗨️ 0 ♺ 0 🤍 17