ฝัง
- เผยแพร่เมื่อ 5 ก.ค. 2024
- a short video explaining how we can install Kohya ss on Runpod for LoRA training for stable diffusion in couple of minutes and start training
00:00:00 introduction about Runpod and this video
00:00:27 selection of the pod and deploying a sample pod
00:01:27 installing Kohya ss for LoRA training on Runpod
00:02:17 getting shareable link for training to run as if it is locally running
00:02:43 uploading files and creating folder structure on the pod in Jupyter lab
00:03:31 start training a LoRA on Runpod quick example
#kohya #kohyass #sd #sdxl #runpod #ai #stablediffusion #a1111
runpod community cloud service
www.runpod.io/console/gpu-cloud
Kohya ss gitbub page with installation instructions
github.com/bmaltais/kohya_ss#...
Command 1:
cd /workspace
git clone github.com/bmaltais/kohya_ss.git
Command2:
cd kohya_ss
./setup-runpod.sh
Command3:
./gui.sh --share --headless
SDXL 1.0 file
huggingface.co/stabilityai/st... - วิทยาศาสตร์และเทคโนโลยี
how do you open tensorboard on the kohya gui. It creates a local link url instead of a public one.
sorry no idea, only the GUI is easily obtained with gradio live link as seen in the video, for tensorboard alternatives might need to be applied such as those suggested on stackoverflow.com/questions/37987839/how-can-i-run-tensorboard-on-a-remote-server/59887887#59887887
Hey! I tried everything how you explain it, but when I press the button "Start training" I geet "image folder path is missing". What am I doing wrong? My path is named "/workspace/data/img/40_myname man" and I made sure that the name matches the files on runpod. I even used "Copy path" to avoid any typing errors. Thanks a lot for your advice.
yes, you should only include the path up to image like this:/workspace/data/img
the training images folder name must not be included only the parent folder
same applies for the class folder if you used regularization, we only include up to /class ... and not inlude the 1_man or 1_person in the paths in Kohya ... these are read automatically by Kohya in folder structure .... also incase you got an error after doing this fix ... and Kohya successfully loads up the number of images you got in each of the training folders, then no worries *double check how many images are being loaded to make sure it is correct during training*
@@AI-HowTo Thank you so much for your help! It worked now. I wish you good luck with your channel. Your Guides are superb. Best regards from Germany.
Thank you so much for your kind words, best of luck to you as well.
Runpod is a genuine disappointment, a theft... I tried it, and luckily, I didn't invest much money in it. Running a virtual machine felt more like playing a slot machine of fortune. One moment it works, you turn it off to try other configurations, and the next moment, you no longer have GPUs available, so you have to set up another machine, paying, of course. I don't like gambling; I despise it. It's a theft, a disappointment. Don't waste your money there. $10 just to train one LORA... thieves.
Do you have to do this every time you want to load/run Kohya on runpod?
short answer is no: some pods have Kohya and other stable diffusion stuff pre installed, but might not always be available or be more expensive...
on this pod, it was a cheap pod, i did install it manually, back then, there was no other pod that offers this service as part of the pod back then... once you install it, it stays, until you delete the pod, and you better delete it if you dont plan to use it daily, to avoid daily storage charges.
@@AI-HowTo Thanks for the speedy reply!
Where is the UI where i can enter the prompts after training?
sorry, this is only for Kohya training installation on runpod (because SDXL training does work on lower VRAM computers), not for A1111 installation and prompting, Runpod has another template called Run diffusion which has A1111 installed and can be used for prompting
When i use kohya_ss and run image captioning, or training or anything -> i don't get any response from anything -> if i in runpod look at logs, there is nothing, like the button pressed are hitting the VOID :) HALP. links, sources, guides anything plz
sorry, never had this problem before, watch the video steps again, hopefully you can figure out what went wrong... if things didnt go well, then try looking up another Pod which has stable diffusion/Kohya already installed... at the time of preparing this video, cheaper Pods didnt have Kohya, so installing it like this was optimal for me.
I have the same issue. Note that this is about the captioning function, which I dont think you use in this tutorial. I am not sure why it is not working... do you have an idea?@@AI-HowTo
Can you make the tutorial for google colab?
sorry, didnt use colab thus far, will check later if I could present something there as well.
Price of the GPU instances has gone up 25% in 3 weeks lol
true, many GPUs have got increase, life changes by the day these days, its really crazy, it is becoming like a rat race :)
Runpod is a genuine disappointment, a theft... I tried it, and luckily, I didn't invest much money in it. Running a virtual machine felt more like playing a slot machine of fortune. One moment it works, you turn it off to try other configurations, and the next moment, you no longer have GPUs available, so you have to set up another machine, paying, of course. I don't like gambling; I despise it. It's a theft, a disappointment. Don't waste your money there. $10 just to train one LORA... thieves.
Runpod is a genuine disappointment, a theft... I tried it, and luckily, I didn't invest much money in it. Running a virtual machine felt more like playing a slot machine of fortune. One moment it works, you turn it off to try other configurations, and the next moment, you no longer have GPUs available, so you have to set up another machine, paying, of course. I don't like gambling; I despise it. It's a theft, a disappointment. Don't waste your money there. $10 just to train one LORA... thieves.
one must indeed by very careful about pricing and usage, prepare everything locally, then only do the final training there to avoid losing money ... I found it the lesser evil to avoid buying a very expensive GPU at the time being....I found most online options are at least as expensive as this option or more, not sure if there is a better service, in my tests, i did couple of LoRAs with less than $4 in total, but it is easy to spend lots of money there if the training results needed repeating and needed higher quality results, not sure what are your choices here, hopefully one can buy his own gear soon that is effective for these AI tests.