I've been doing ML computer vision stuff mostly as a hobby the past two years, mostly stuff like object detection, bounding, and segmentation , and just stuff that help me organize my pictures. what should I be learning to go from being a dilettante to someone proficient in the field? I heard some tell me to learn the math behind it , I'm little bit apprehensive about that since as someone that did an EE program nearly all the theoretical math i went over in there never came up again, is ML different in this respect.
Noah Hill
the hell? what is the script using the decode the image?
Leo Russell
What do you mean?
Jonathan Parker
oh yeah meant here , got confused
Ayden Edwards
whats the memory requirements for ESRGAN? I keep getting OOM when I try it
Interesting news. Paper from Google that shows a technique better than textual inversion for getting consistent characters and objects in a different context dreambooth.github.io/ you could fine tune the model with your own pictures of a subject and it will keep perfect fidelity in new contexts. This will solve the weakness the current AI have. > "How does this differ technically from Textual Inversion?" > "One key component of our method is to finetune using the images of a subject, different from Textual Inversion. This allows for strong identity preservation which is super important for this problem. You can also check the Textual Inversion first author's response to this question here. twitter.com/RinonGal/status/1563092813310410752?s=20&t=N1nlP3JVrnLyVwN2osvOoA We are working on code and next steps. Our method is general and should work on other text-to-image models that are not Imagen, and we would like to implement it on some of the latest open-source models!"
I still don't know what you mean. I tried to upload a 2048x2048 pic but it was too big so I resized it.
Adam Thomas
>did you think it was just equivalent to MS Paint resize?
No and I wouldn't dare to be that much of an arrogant but it is obvious that I lack basics of how it really works, so thank you for informative post and Wikipedia link. I was comparing its results to Gigapixel's so there's my confusion about image manipulation levels.
Ryan Johnson
What CSS is that and how did you add it to the webui? Also, sadly, in the eyes of AI the US of A is still the first world.
My small-ish company does a shitload of ML although I myself am not involved with that side of it, from what I understand it's most important to be smart, comfortable dipping into C++, and at least capable of understanding the maths if a task requires that you really dive into that. I'm guessing that if you were to spend 200 hours across 3 months training in something, though, more of the same of what you've done would be more valuable for finding a job and being useful in it than 200 hours of the underlying maths. OTOH reading your way through one respected underlying textbook in that time would probably be good just to prove you're capable of understanding the theory, so that you can go apply it when needed. Maybe someone who actually works in the field can give you better advice than my guesses though.
Dylan Barnes
>try using fix faces with gfpgan >error pops up saying it can't load C++ ops >both versions of pytorch and torchvision are compatible wat do?
found a bug: when using batch, seed for the first output gets written in the yml for all outputs so all the filenames I've been giving to the images have been bs
Just a guess, but do you have two virtual envs with the same name?
Kayden Kelly
only joshing with you mate. Happy friday
looks sick although google didn't release imagen even to API and I can't see any concrete claim that they'll actually release this, rather than just "get it working with" SD/DALLE but keep it for nobody or academics only. Maybe I'm blind
Austin Green
>. Maybe someone who actually works in the field can give you better advice than my guesses though. I applied for CV jobs knowing I couldn't actually get them just so that I can get a phone interview so I can ask them directly, they only gave me very general/vague and unhelpful answers. Maybey its a dick move to waste their time like but I dont know who else to ask, I want to learn I just dont know what direction to go.
Justin Thompson
is installing the SDKs actually required/recommended? I do not have them installed and have been running without issue
Jaxson Reed
Hey can anyone help me generate giantess content
Ai sucks at making giga-giantess stuff
Alexander Sullivan
Arch and 6700xt
James Reed
reinstall visual studio and cuda
Brody Gray
from last thread im currently generating more avatar/navi images
I don't know for sure what GUI that is but that's not the behaviour in Webui, however that's because it manually increments the seed by 1 and does another pic, up to the "batch size". i think the original script from SD however takes your batch size and generates pics one by one from that seed, so like, seed 42069 generates a certain pic and the 2nd seed from it is unreachable except by generating from 42069 again and generating at least 2 pics. that's why webui changed it, if you make a dozen pics and the 11th one is great you have to at least generate another 11 to work with it again
I'm using hlky's webui from the guide it getsthe seed right in the file name but wrong in the yaml
bs=bullshit
Aaron Scott
>パンツ Garbage.
Disclaimer: Though I do have (very limited) academic experience with computer vision this is not my main field of study. I will talk about neural networks and statistics more broadly.
Practical side: Stanford University has published freely available courses for neural networks. The exercises have you implement standard feed-forward neural networks using NumPy routines. The exercises are honestly pretty hard but they do a good job of teaching you how to implement efficient numerical computations with NumPy/TensorFlow (NumPy routines are literally 100x faster than Python loops). Requires a lot of linear algebra.
Theoretical side: Learn about the method of maximum likelihood. This is the foundation for cost functions such as cross entropy or mean squared error. An important point is how bias and variance is defined in this context and how it relates to neural networks. Statistics is a very deep rabbit hole though.
Quite honestly though, the field of "data scientists" and ML practicioners is already full of dilettantes. If you can name the correct definition of linear regression and explain how it applies to xy data you will already be in the top 10%.
>100x faster than Python loops thats that one thing I noticed when I starting learning python, loops are fucking slow in this, why is all machine learning scripted in it?