Arcgis learn with RTX 3090

williamlidberg1

I have successfully trained and used models in arcgis pro 2.6 using  a gtx 1050 ti. I recently upgraded to a RTX 3090 to scale up my project. However RTX 3090 require cuda 11 which is not supported by pytorch 1.4 that comes with arcgis pro 2.6 or the deep learning package. Is there a plan to update the arcgis learn package to work with pytorch 1.7 or cuda 11? Can I solve this on my own?

JoshuaBixby

ArcGIS Pro doesn't come with PyTorch, so this is really more of Deep Learning issue than a general Pro issue.

Looking at Deep Learning Essentials :: Anaconda Cloud, which currently has the pre-release Pro 2.7 deep learning packages, it appears deep learning has a specific requirement not to use cudatoolkit 11.

{
  "arch": "x86_64",
  "build": "arcgispro_6",
  "build_number": 6,
  "depends": [
    "arcgispro 2.7.*",
    "cudatoolkit >=10.1,<11",
    "cudnn >=7.6,<8.0a0",
    ...
    "pytorch 1.4.*",
     ...
  ],
  "features": "arcgispro",
  "license": "Apache 2.0",
  "name": "deep-learning-essentials",
  "platform": "win",
  "subdir": "win-64",
  "timestamp": 1605904242370,
  "version": "2.7"
}

 

Getting back to your questions....  Is there a plan to have deep learning tools work with cuda 11, I think the answer is most definitely.  Does that plan happen with Pro 2.7, it doesn't look like it.  Given the way the configuration files look on Anaconda cloud, it appears you would not be able to update some of the dependency packages yourself.