Code is adapted from this the colab notebook. Thanks to the contribution from community.
Add alternative train_dreambooth.py from this colab notebook. use get_alt_script.ps1 to download it.
Choose PyTorch 1.11.0 Python 3.8 (Ubuntu 20.04) as base image.
cd ~
git clone https://github.com/crosstyan/dreambooth-scripts-for-autodl dreambooth
cd dreambooth
git submodule update --init --recursive
# TODO write an init script to help configure the envrionment
# for now you have to do it manuallyusing conda/mamba with Python 3.10.6.
First of all you have to install conda and mamba. I assume you have done that.
# See https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-python.html
mamba create -n diffusers python=3.10.6
# if you have installed powershellIf you're NOT using my scripts you can choose to configure bash
conda init bash
conda activate diffusers
# or use your favourite text editor
# or sed, whatever
# see also https://stackoverflow.com/questions/17701989/how-do-i-append-text-to-a-file
vim ~/.bashrc
# echo "conda activate diffusers" >> ~/.bashrc
# add `conda activate diffusers` to the end of fileIf you're using my script you have to install PowerShell. See Installing PowerShell on Ubuntu.
# after installing PowerShell
conda init powershell
conda activate diffusers
vim ~/.config/powershell/profile.ps1
# add `conda activate diffusers` to the end of fileWhy do I use PowerShell? Because I can't write correct bash scripts (Help wanted!) and too lazy to use python.
I'm using the diffusers fork of CCRcmcpe, which added wandb support and a few improvements.
git clone https://github.com/CCRcmcpe/diffusers repos/diffusersLLVM10 is required as well
# https://packages.ubuntu.com/focal/llvm-10-dev
apt install llvm-10-dev# install these packages
# switch to TUNA PiPy mirror if you get any error
# https://mirrors.tuna.tsinghua.edu.cn/help/pypi/
pip install -U pip
pip install wandb
pip install -U --pre triton
pip install accelerate==0.12.0 transformers ftfy bitsandbytes gradio
pip install omegaconf einops pytorch_lightning
pip install transformers
cd repos/diffusers
pip install .xformers is totally OPTIONAL. You can skip this part if you feel like doing it. It just speed up the training process, which is trivial if you have a beefy machine like A5000. If you mess up anything about xformers, just uninstall it.
pip uninstall xformers
prebuilt wheel build with
Cuda compilation tools, release 11.3, V11.3.109
RTX A5000
Python: 3.10.6
OS: Ubuntu 20.04.4 LTS x86_64
Kernel: 5.4.0-100-generic wget https://github.com/crosstyan/dreambooth-scripts-for-autodl/releases/download/v0.0.14/xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl
pip install xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whlHere's how you build it from source.
# https://github.com/C43H66N12O12S2/stable-diffusion-webui/releases/download/linux/xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl
# won't work since AutoDL provided Ubuntu version is too old
# GLIBC_2.32 is required
git clone https://github.com/facebookresearch/xformers repos/xformers
cd repos/xformers
# install ninja to speedup building.
pip install ninja
# or maybe
# apt install ninja-build build-essential
pip install -r requirements.txt
# use `pip wheel .` to create a whl file
pip install .I'm not sure if
MAKEFLAGS is effective since it still takes a long time tocompile and still only one core be used. I mean a about an hour or less, not sure.
Using ninja could speed the building process up. (source)
OPTIONAL! Useful if you want to add tags to your images.
pip install -r repos/DeepDanbooru/requirements.txt
pip install basicsr
pip install -U numpy # I have to do this
pip install repos/DeepDanbooruSee Usage
- Provide a Jupyter interface directly
- Intergrate AUTOMATIC WebUI (I'm afraid there's no enough space)
Error in File "/root/miniconda3/envs/diffusers/lib/python3.10/site-packages/bitsandbytes/cuda_setup/paths.py", line 90 that CUDASetup.get_instance is a function.
Somehow the get_instance function is not called. Add () to fix it.
Is the TUNA still caching the old version? This kind of bug should be fixed already. (commit)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent