Replies: 7 comments 2 replies
|
....ok i'll take al ook :) :D and yea Claude is the model by Anthropic that i pay to use :D |
|
Also next time dont bundle an exe file, virus total is flipping out at me currently and defender hasn't even seen the file yet XD I'll check your python file tho in notepad! |
|
Ok my claude code and I worked through your idea and this is our review -- please don't take this as i know everything or that i'm knocking you lol: Hey! First off — the AI-calculated parameter optimization is genuinely cool work, especially as a starting point. That's more than most people attempt. A few thoughts from someone who's been deep in the weeds building a UI/training system on top of sd-scripts: On the architecture: You'd get a lot more mileage building your calculator/automation layer to output configs for existing trainers (sd-scripts, etc.) rather than bundling your own training loop. The number of variables in the Windows ML environment alone — CUDA versions, driver states, Python environments, VRAM edge cases — is brutal. Wrapping a proven trainer means you only have to solve the automation problem, not the automation AND the training problem at the same time. On "optimization": What you've built is really a parameter suggester based on heuristics, which is valuable! But true optimization requires a feedback loop — train, evaluate outputs, adjust, retrain. Without that closed loop, calculated values are starting points, not optimized results. The dirty secret of LoRA training is that the "optimal" values for one dataset can be completely wrong for another with identical specs. That's why communities share presets based on thousands of real training runs rather than formulas. On scaling: The more features you add (inference, advanced scheduling, custom optimizers), the more every variable compounds. Talking to existing trainers headlessly keeps your scope manageable and lets you focus on what's actually novel about your project — the intelligent parameter selection. None of this is meant to discourage — it's meant to save you from the same rabbit holes we've already fallen into. The calculator idea has legs, it just needs to be pointed at the right target. |
|
Right now, TOS is set to train at the very high end of the spectrum, almost to the point of overfit. Thinking of making those parameters user accessible. to allow for personal fine tunning.
I'm not building my own trainer. I am going to use yours as soon as I can figure it out. Cause is Kohya based and my experience with that is minimal. So basically, what I tried to accomplish is: if I can establish the optimal exposure each individual image in my dataset requires based on community guidelines, then I can move the levers to achieve said exposure. Did you noticed calculations change depending on Lora type? so I know "The dirty secret of LoRA training is that the "optimal" values for one dataset can be completely wrong for another with identical specs. " is absolutely true. I already trained a couple of lora's using this method and they appear to be functional. I have about 20 loras in the Training Queue either completed or pending completion. The time to completion numbers under AI-toolkit are real world measurements of real Lora training sessions. those loras are my experimental sample. I've been working on this idea for a bit. It all started in an excel sheet 😁. Let me know if you can train one of yours via this method for an impartial assessment. whenever you can, if you can. |
|
So i'm NOT saying no here LMAO I used claude to peek at most of what ws up and i am NOT that good at math I approach training by "FAFO" = "Whatever works most of the time" My CAME preset i made is based on Novowels, because character not having the base style = REALLY GOOD 99% OF THE TIME (This does't always work btw, xD illustrious is a bitch) The problem with me though is i'm repetetive in how i train, because i don't understand the math logic - You're WELCOME ALSO TO HELP DEVISE more presets for this shit too, so based on your experiments you literally could use your automation and otherwise to figure the math out.. Dgracey my dude, honestly I SUCK AT MATH, i designed the system around the fact i don't do well with logic XD that it is DESIGNED more in tune with the idea that sd-scripts as a CLI is a nightmare, and Bmaltais in gradio follows that sort of ML pattern.. I used Next/Node because i treat AI like art. I am like. I HAVENT EVEN RE TESTE THE SYSTEM YET XD And i want to BUILD A QUEUE FOR THE TRAINER but yea -- IF you can come up with MATH AND LOGIC FOR IDEAS FOR THE TRAINER even if you're not building your own: YOU CLEARLY KNOW MATH: Math + LOGIC = I go nuts i dont understand it all half the time lol. But yea you're welcome to come up wiht more ideas -- this one you'll have to help me understand a bit more because ven claude was over my head this time but im' tired right now XD I tried to sit and understand Final Fantasy 7's storyline.. |
|
Who me? math!? nope! I only have ideas that either Claude or ChatGPT helped me make it into numbers. wait until you see what I've been working on lately 😁 |
|
Yea but you seem to have more of an idea of LOGIC than I do lol, that's more what i'm saying - you have a grasp of what CONNECTIONS there is in the "MATH" end -- I'd look at the wires and go "Cut the whatever one i dunno" -- "I failed flow charts" XD but yes, in reality I'm TECH ADJACENT as i learned a LITTLE bit about programming, and used to do web design XD And here i am back in the saddle learning it all over again. And getting hell on threads for it. I haven't had time to re-test the system again cause i'm exhausted i'm walking baxter five times a day and while it doesn't seem like much it's like 15-20 minutes each time - and i'm also doing other extra things around the house while my stepdad's in a cast XD |
|
MY BRAINS JUST REALIZED WTF YOUR ASKING ON THIS LMAO. Ok i dunno. Like my brain is never entirely here so i THINK IN SOME WAYS THIS REALLY IS A DECENT IDEA because i'm thinking abotu the ide athat the STEP CALCULATOR NEEDS MORE WORK ANYWAYS Paramaterization of hte step calculator - "What might work for you doesnt work for everyoen" right now it's just a basic calculator based on a simple equation from KohyaSS. But it doesn't factor in batch size, which GPU you have, it doesn't even factor in that you CAN train on less than 1000 steps on CAME -- it's bias is equal to the original equation, it's not the AI's fault that the calculation is just .. old XD I peeked at what you're making finally and by god as stated like before I'm tempted to rope you into the team to dream up stuff cause middle aged nerds unite XD |
Uh oh!
There was an error while loading. Please reload this page.
@duskfallcrew Seeing you (all?) work thru such a complex wall of code to get this thing running inspired me to learn. Thanks for introducing me to Claude! such a nice...guy? gal?...AI? anyway.... I would like to offer an idea of mine for consideration: I've been working on a spreadsheet to facilitate and, to a point, automate LoRA training. My idea is based on a concept I call "TOS" as in Target Optimization Signal. It establishes a baseline signal per image. Based on LoRA type. So, I would like to humbly submit my paper for peer review.
Lora_Calculator.zip
All reactions