Installation issue about omni3d

[genevieve-tech]

I also encountered similar issues with Detectron2 and Pytorch3D, and used the following commands to install them:

• python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
• pip install 'git+https://github.com/facebookresearch/pytorch3d.git'

However, there still seem to be incompatibility issues when I run demo.py.

---

Edit: after several reinstallations, I was able to run demo.py with the commands provided in the README. Not sure what happened exactly.

[matteosodano]

I think the problem comes from the installation instructions you give. We executed them weeks ago, when you first published the repo (WITHOUT the training code), and the versions of some libraries were different from the ones we get now executing the exact same commands (pytorch3d becomes 0.7.0 from 0.6.2, detectron2 becomes 0.6 from 0.4). In fact, the old conda env is able to start training, while the new one isn't.

My suggestion would be to publish directly a requirements.yml file with the exact versions you used for training.

[gkioxari]

Hi @matteosodano. It seems that you weren't able to install detectron2 properly.

1. Are you using conda and are you following the commands we provided?
2. What is your cuda version
3. Our code should be compatible with newest versions of pytorch3d, detectron2 and pytorch. There should be no incompatibility issues with newest versions. I think the issue that you are facing is that your environment was not able to install detectron2 properly. So let's figure out why that is.

[matteosodano]

Hi

1. Yes, I'm using conda with the commands you provided
2. I tried on two machines: one with cuda 11.2, another with cuda 10.2
   2a. On that same machine with cuda 10.2, I was able to install everything properly and run the demo some weeks ago.

[gkioxari]

What were you not able to install with cuda 11?

[matteosodano]

It looks like the problem, as you suggested is detectron. All the rest gets installed succesfully.