iTranslated by AI
How to Run kohya-ss LoRA on Paperspace
Introduction
If you start Paperspace from this link, you will receive $10 in credit.
I followed this article.
I made some changes to the code for my own use.
I wrote this for record-keeping purposes as a memorandum, so some parts may be difficult to understand.
The setup and startup of Paperspace are explained with images on the reference site, so it may be easier to understand there.
Subscribing to PaperSpace Pro
It is a bit confusing, but there are pay-as-you-go and monthly plans.
Pay-as-you-go can be used on an hourly basis.
If you subscribe to a monthly plan, "free" will be displayed in the GPU section.
Pro is available for $8 per month. You can use GPUs with up to 24GB of VRAM as many times as you want in 6-hour increments.
*Note: You may not be able to use a GPU if there is no availability.
Here is a quick reference table for GPUs and plans.
Source: https://qiita.com/kunishou/items/dccb44848e5b572619bc
If you start Paperspace from this link, you will receive $10 in credit.
Project Creation and Startup
Log in to Paperspace and proceed by referring to the following site.
We will proceed using Gradient.
Please be careful not to proceed with Core by mistake.
Installation of LoRA
WARNING: Running pip as the 'root' user can result in broken permissions and
conflicting behaviour with the system package manager. It is recommended to
use a virtual environment instead: https://pip.pypa.io/warnings/venv
You don't need to worry about this message appearing during pip.
mkdir LoRA
cd LoRA
sudo apt update -y && sudo apt upgrade -y
git clone https://github.com/kohya-ss/sd-scripts.git
git clone https://github.com/derrian-distro/LoRA_Easy_Training_Scripts.git
cd sd-scripts
apt -y install python3.10
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
pip install -U -r requirements.txt
pip install -U --pre triton
pip install -U -I --no-deps xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl
// This part could not be installed due to an OSError. If your goal is LoRA training, you might not need to do this.
cd ..
accelerate config
//- This machine (enter)
//- No distributed training (enter)
//- NO (input)
//- NO
//- NO
//- all
//- fp16(enter)
cp LoRA_Easy_Training_Scripts/lora_train_command_line.py sd-scripts
cd sd-scripts
This completes the environment setup. All that's left is to run LoRA.
You can execute it with accelerate launch.
I run it as follows.
Please replace the "[file path]" parts.
accelerate launch --num_cpu_threads_per_process 12 train_network.py --pretrained_model_name_or_path="[file_path.ckpt]" --train_data_dir="[file_path]" --reg_data_dir="[file_path]" --output_dir="[file_path]"--resolution=320,960 --train_batch_size=4 --learning_rate=8e-5 --max_train_epochs=10 --save_every_n_epochs=1 --save_model_as=safetensors --clip_skip=2 --seed=42 --color_aug --network_module=networks.lora --keep_tokens=7 --enable_bucket
In the above case, there are four parts you need to change yourself.
-
--pretrained_model_name_or_path=
This is the file path to the .ckpt file of the model you want to perform additional training on.
Example:
--pretrained_model_name_or_path=/notebooks/stable-diffusion-webui/models/Stable-diffusion/Evt_M.ckpt
-
--train_data_dir=
Specifies the folder for the training materials.
Example:
--train_data_dir=/notebooks/LoRA/Training
There are naming conventions for folder names, so please refer to this for details.
-
--reg_data_dir=
Specifies the folder for regularization images.
Example:
--reg_data_dir=/notebooks/LoRA/reg
-
--output_dir=
Specifies the folder where the training results will be saved.
Example:
--output_dir=/notebooks/LoRA/testLoRA
Startup from the Second Time Onwards
cd LoRA
sudo apt update -y && sudo apt upgrade -y
cd sd-scripts
apt -y install python3.10
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
pip install -U -r requirements.txt
pip install -U --pre triton
pip install -U -I --no-deps xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl
// This part could not be installed due to an OSError. If your goal is LoRA training, you might not need to do this.
cd ..
accelerate config
cd sd-scripts
Recommended Books for Image Generation
[2023 Latest] Includes over 1,000 types of prompts! Spellbook for NovelAI and Local Use: Tips, Techniques, and Recommended Tools (Introduction for AI Artists) Kindle Edition
Becoming a God-tier Artist through AI Collaboration: Understanding Stable Diffusion from Research Papers
Leading-edge Technology: How to Use AUTOMATIC1111 / Stable Diffusion web UI
Discussion