Pytorch3d: Difficulty in installation

[yyl494]

Here is some error in the 2nd step of installation, which is installing prebuilt binaries for Pytorch3d. But I think my core library is correct.

Here is my conda list. torch version is 1.4.0

my python version is 3.6 and cuda 10.0 in 'nvcc -V'. I don't know what is '__cuda = 10.2' is in the following picture.

[bottler]

You don't have pytorch (or torchvision) installed in the list. That's why it's complaining.

[yyl494]

You don't have pytorch (or torchvision) installed in the list. That's why it's complaining.

But I can see my toch and torchvision in the last picture in conda list. I can import them in this environment

[bottler]

Sorry. I missed them. Something funny is happening. It looks like torch is installed from pypi not conda - is that right? That means you don't actually have pytorch conda package.

Can you maybe try in a new repository, just installing what you need.

It is true that we don't support cuda 10.2, but I don't think that's a problem here, you have the 10.0 cudatoolkt.

[yyl494]

I use pypi because of a firewall. Can I use torch 1.3 from a mirror website？

[bottler]

If you are happy with CPU-only for some parts of pytorch3d, you can try pip install pytorch3d in that environment.

Another possibility is building pytorch3d from source in that environment.

[yyl494]

Thank you! I'll try these.

[yyl494]

I solve the problem with conda install --offline for torch, torchvision and pytorch3d. It's really important to use conda and pypi is the reason of conflicts. Thank you.