Installation and Setup with DockerFile

[Hemantr05]

🚀 Feature

I would like to propose and implement a DockerFile for easy installation and set up of pytorch3d

Motivation

A lot of users might face problems while installation, I personally took some time to set it up, but a dockerfile makes this easier.
One line of code to run the dockerfile.

Pitch

The DoclkerFile would :

1. check for python
2. Ask for OS
3. create a virtual environment
4. Install all requirements
5. Checks for cuda (and version) if doesn't exist, will install with pip (without cuda support)
6. User can choose.
   [a] anaconda [b] pip
7. User can choose building / installing from source.
   [a] Install from Github [b] Install from local clone

NOTE: we only consider adding new features if they are useful for many users.

[nikhilaravi]

@Hemantr05 yes this sounds like a good idea. Do you plan to implement it or do you need support from us?

[Hemantr05]

@nikhilaravi will try implementing it. Should be done by today.
If any support is needed, what would be the approach for it?

[nikhilaravi]

Great! Feel free to make a PR when you're ready. If you have any questions you can ask them here or on your PR.

[Hemantr05]

Will do, thanks a ton @nikhilaravi

[bhadresh74]

I was successfully able to setup pytorch3d using docker. Let me know if anyone needs any help.

[Hemantr05]

@bhadresh74 is it possible for you share the same?

[bhadresh74]

install_Pytorch3D_Dockerfile.txt
@Hemantr05 There it is. Enjoy.

[Hemantr05]

@bhadresh74 thank you

[gkioxari]

Seems like this issue is resolved. I am closing it and feel free to reopen it if anything new emerges.

[tianyilt]

It seems not to work today because pytoch3d=0.2.0=py36_cu100_pyt14 has removed in -c pytorch3d-nightly
The solution is to replace RUN conda install --name pytorch3d pytorch3d -c pytorch3d-nightly with RUN conda install --name pytorch3d pytorch3d -c pytorch3d because pytoch3d=0.2.0=py36_cu100_pyt14 exists in -c pytorch3d

[ktiwary2]

How do you upgrade the pytorch3d version after running the docker? Running conda update doesn't seem to work (im running it inside the conda base env).

[willhama]

install_Pytorch3D_Dockerfile.txt @Hemantr05 There it is. Enjoy.

Dear @bhadresh74, within your dockerfile you are using python 3.6, but the recommended one is 3.8. I am trying since hours to install it with 3.8 but it doesnt work. Do you mind helping me on that?

[abhayraw1]

An official docker image is really needed to avoid the countless hours being spent by researchers and developers trying to get it installed!

[RujingD]

Please,How can I use pytorch3d in docker?

[domattioli]

An official docker image is really needed to avoid the countless hours being spent by researchers and developers trying to get it installed!

This x100 for Windows users.

[bhezy]

@bhadresh74 is there a way to make a Docker image that generates pytorch3d v0.7.0, for use on a Windows machine? I have been trying, but it seems as though this Docker File is not working (I am on Windows, and have not made any changes). I assume that it should work regardless of machine, I have just been having a lot of trouble getting things to work.

@bottler, please consider an official Docker image for those of us in the community on Windows machines. It is extremely difficult the build pytorch3d on the Windows OS, which is unfortunate given how important it is in the realm of research. I am certain that the team who developed it can come up with a phenomenal solution.

[bottler]

If you want to use PyTorch3D in docker rather than on windows itself, you can use an existing package; I don't see why a pytorch3d specific image would be that useful. For example, starting with an official pytorch package, these conda install steps are fast:

docker run --rm -it pytorch/pytorch:1.11.0-cuda11.3-cudnn8-runtime
> conda install -c fvcore -c iopath -c conda-forge fvcore iopath
> conda install pytorch3d=0.7.0 -c pytorch3d
> ipython
>> from pytorch3d.structures import Meshes