Added docker & docker-compose to the project

[rowellz]

Added support for docker-compose. This docker container will enable users to run both acezero & nerfstudio, allowing the creation of gaussian splats from just a few commands. I haven't fully tested all of the features of acezero, but am aware that some functionality like --render_visualization doesn't work in the docker container.

Only added two files, the Dockerfile & docker-compose.yml file, and updated two files, the .gitignore & README

Hoping this will make ace0 more accessible and easier to contribute on!

Also, this was tested on a RTX 3060 12GB card on a Ubuntu host machine. Was able to successfully train a gaussian splat from datasets of 6000 & even 10000 images, both of which matched roughly 70% of each dataset and the camera calibration results were quite impressive based on the nerfstudio splat. I'm going to try this on a larger gpu so I can hopefully train the splat at a higher resolution, since my 3060 can only handle a downscale value of 6.

[amonszpart]

Hello, thanks for your contribution, and effort to try to make ACE0 easier to use!

I left a few optional suggestions, I hope you consider them.

Cheers,
Aron

[amonszpart]

Would you maybe be able to consider using a free conda implementation? E.g. Miniforge from conda-forge?

It's hard to predict what will happen to the permissiveness of Anaconda products: https://www.theregister.com/2024/08/08/anaconda_puts_the_squeeze_on/

If not, please add an echo or some sort of warning for users to check Anaconda's licensing terms.
Thanks!

[amonszpart]

Suggestion: Could you please combine as many apt-get calls as possible to a single RUN layer? See also https://docs.docker.com/build/building/best-practices/#apt-get

In general, it's good to have fewer RUN lines, and group instructions with && \ at the end. This allows the layer to not contain changes that would be undone in a consecutive command, and will result in smaller Docker images that are easier to carry around.

[amonszpart]

I think all these RUNs up to L40 could be combined (or if not, very few splits would be needed). See advice from the Docker blog:

At a high level, it works by analyzing the layers of a Docker image. With every layer you add, more space will be taken up by the image. Or you can say each line in the Dockerfile (like a separate RUN instruction) adds a new layer to your image.
(source: https://www.docker.com/blog/reduce-your-image-size-with-the-dive-in-docker-extension/)

RUN apt-get install -y libglib2.0-0  # TODO: merge with earlier ones

ENV TCNN_CUDA_ARCHITECTURES="50;52;60;61;70;75;80;86" # consider making an export if it's not needed for the final image

RUN && \
  conda create --name nerfstudio -y python=3.8 && \
  conda run -n nerfstudio pip install --upgrade pip && \
  conda run -n nerfstudio pip install "torch==2.1.2+cu118" "torchvision==0.16.2+cu118" --extra-index-url https://download.pytorch.org/whl/cu118 && \
  conda run -n nerfstudio conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit && \
  conda run -n nerfstudio pip install ninja "git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch" && \
  conda run -n nerfstudio pip install nerfstudio && \
  conda run -n nerfstudio ns-install-cli

[amonszpart]

suggestion: make this the default "7007:7007" or closer to it, e.g. "7008:7007"

[amonszpart]

Suggestion: put a comment above explaining what this does and why it's needed.

[amonszpart]

Suggestion: remove this, it's not needed for the final container (unless I'm mistaken)

[amonszpart]

I usually use docker run --gpus='all,"capabilities=compute,utility,graphics"' ..., now that might not be the most up to date way, but I think capabilities=all would be nice to have, see also here: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/docker-specialized.html#driver-capabilities

[rowellz]

I think the docker-compose docs recommend using capabilities: [gpu] for docker-compose files

https://docs.docker.com/compose/how-tos/gpu-support/

[rowellz]

@amonszpart Thanks for reviewing this! I almost forgot about this PR, haha. I'll try to resolve most of if not all of your feedback by the end of this week, my apologies as I have other priorities at the moment!

[rowellz]

@amonszpart Hey! I think I took care of most if not all of your feedback, l did forget to mention that the extra_hosts param allows the nerfstudio client to be accessed by the host machine's IP address & localhost during the Gaussian splat training phase. Just let me know if you want anything else tweaked so it's good to merge!

Cheers 👍