Docker support (Tested on RTX3080) + ignoring reproducable files / cache

.dockerignore

@@ -0,0 +1,6 @@
+models/
+Dockerfile
+README.md
+LICENSE
+assets/
+*.egg-info

.gitignore

@@ -0,0 +1,5 @@
+models/ldm/text2img-large/
+outputs/
+src/
+__pycache__/
+*.egg-info

Dockerfile

@@ -0,0 +1,27 @@
+FROM nvcr.io/nvidia/pytorch:22.04-py3
+MAINTAINER Peter Willemsen <[email protected]>
+
+RUN conda
+
+RUN mkdir -p /opt/ldm_package
+COPY ./setup.py /opt/ldm_package
+COPY ./ldm /opt/ldm_package/ldm
+COPY ./configs /opt/ldm_package/configs
+COPY environment.yaml /opt/ldm_package
+
+# For the ldm-dev user
+RUN chmod 777 -R /opt/ldm_package
+
+WORKDIR /opt/ldm
+
+# Add dev user
+RUN useradd -ms /bin/bash ldm-dev && \
+    usermod -aG sudo ldm-dev && \
+    echo '%sudo ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers
+USER ldm-dev
+
+RUN conda env create -f /opt/ldm_package/environment.yaml
+RUN pip3 install -e /opt/ldm_package
+RUN conda run -n ldm pip install pytorch-lightning==1.5
+
+ENTRYPOINT ["conda", "run", "-n", "ldm"]

README.md

@@ -62,6 +62,19 @@ Furthermore, increasing `ddim_steps` generally also gives higher quality samples
 Fast sampling (i.e. low values of `ddim_steps`) while retaining good quality can be achieved by using `--ddim_eta 0.0`.  
 Faster sampling (i.e. even lower values of `ddim_steps`) while retaining good quality can be achieved by using `--ddim_eta 0.0` and `--plms` (see [Pseudo Numerical Methods for Diffusion Models on Manifolds](https://arxiv.org/abs/2202.09778)).
 
+## Installing on Docker
+
+- Build the image: `docker build . --tag latent-diffusion`
+- For text-to-image, download the pre-trained weights (5.7GB):
+  ```
+  mkdir -p models/ldm/text2img-large/
+  wget -O models/ldm/text2img-large/model.ckpt https://ommer-lab.com/files/latent-diffusion/nitro/txt2img-f8-large/model.ckpt
+  ```
+- Sample with (Make sure to call in the directory of this repo):
+  ```
+  docker run --name=tmp-diffusion --rm --gpus all -it -v "$(pwd):/opt/ldm" latent-diffusion python /opt/ldm/scripts/txt2img.py --prompt "A large blue whale on a freight ship, vector art" --ddim_eta 0.0 --n_samples 4 --n_iter 4 --scale 5.0 --ddim_steps 50
+  ```
+
 #### Beyond 256²
 
 For certain inputs, simply running the model in a convolutional fashion on larger features than it was trained on