Skip to content

Latest commit

 

History

History
63 lines (49 loc) · 3.38 KB

CONDA_SETUP.md

File metadata and controls

63 lines (49 loc) · 3.38 KB
Raw

My original diffusers conda environment setup was:

conda create -n diffusers
conda activate diffusers
conda install -y pytorch torchvision cudatoolkit=11.6 -c pytorch -c conda-forge
pip install diffusers transformers piexif

The module env_config.py specifies the following:

pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
pip install omegaconf==2.1.1 einops==0.3.0 pytorch-lightning==1.4.2 torchmetrics==0.6.0 torchtext==0.2.3 transformers==4.19.2 kornia==0.6
git clone https://github.com/deforum/stable-diffusion
pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
pip install -e git+https://github.com/openai/CLIP.git@main#egg=clip
pip install accelerate ftfy jsonmerge matplotlib resize-right timm torchdiffeq
git clone https://github.com/deforum/k-diffusion/
echo > k-diffusion/k_diffusion/__init__.py

...however on your own system (as opposed to in a Colab notebook) I do not recommend this.

  • For one, it uses pip not conda (so can't take advantage of mkl accelerated numpy etc).
  • Secondly, it uses CUDA 11.3 specifically, whereas 11.6 is available.

I looked into the effect of the final echo (that blanks the init file) and came to the conclusion its purpose is to avoid (perhaps slow) loading of: augmentation, config, evaluation, gns, layers, models modules, (perhaps so that it avoids needing to incorporate the package-level requirement jsonmerge which is imported in config, or clean-fid in evaluation...)

In light of this, it'd still be simpler I think to just git clone it, overwrite the __init__ module, and then install it in editable mode. (I don't imagine those dependencies are troublesome)

If you do this, you get CLIP too, as the last line of the k-diffusion package's requirements is git+https://github.com/openai/CLIP.

Another thing to note is that pip install -e git+GIT_URL#egg=foo is the same as git clone GIT_URL src/foo; pip install -e src/foo.

deforum/stable-diffusion won't let you do this as the metadata has the name latent-diffusion, so you must pass this as the #egg name in the URL param or it'll error out cautiously. The internal name is ldm (as in, the code calls import from ldm.models... import ...)

With this in mind, to recreate an environment:

git clone https://github.com/lmmx/deforum-stable-diffusion
cd deforum-stable-diffusion
conda create -n deforum
conda activate deforum
conda install -y pytorch torchvision torchtext cudatoolkit=11.6 pytorch-lightning -c pytorch -c conda-forge
pip install accelerate einops ftfy jsonmerge matplotlib kornia omegaconf opencv-python resize-right timm torchdiffeq torchmetrics transformers
pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
pip install -e git+https://github.com/deforum/stable-diffusion.git@main#egg=latent-diffusion
pip install -e git+https://github.com/deforum/k-diffusion.git@master#egg=k-diffusion
  • Note that it's important to cd into the package after you clone it so that pip install -e git+https://... will put packages into src/ of this repo, where it can detect the config files (not accessible via Python namespaces).

Having done this, I've turned off the subprocess approach to installation in the code (report_env() takes an arg setup_environment that defaults to False).