How to Install FastAI with GPU support

How do I install FastAI on the Nvidia Jetson-nano (4GB) Development Kit?

I’ve been able to install the provided PyTorch and test whether it detects GPU sucessfully, but I’ve been so far unable to install FastAI.

I’ve tried:

  • pip3 install fastai gets stuck for a long time building the SpaCy wheel and never seems to finish.
  • pip3 install fastai after manually installing some of the requirements, but It removes the manually installed torch and replaces with a non-GPU enabled one. Also crashes when loading on python because the used python version is too old (3.6).
  • using conda, with a more recent python (3.8), but does not use the GPU
  • building fastai from source, but it fails when building fastcore with
ERROR: Cannot install fastcore because these package versions have conflicting dependencies.

The conflict is caused by:
    nbdev 1.2.3 depends on nbconvert>=6.1
    nbdev 1.2.2 depends on nbconvert>=6.1
    nbdev 1.2.1 depends on nbconvert>=6.4.1
    nbdev 1.2.0 depends on nbconvert>=6.4.1

I feel like a big part of the problem here is that the python on the jetson nano is too old, but I don’t know how to build the torch wheel for a more recent version.

Any advice on how to get fastai to work is appreciated, so far I’ve not been able to find a configuration/alternative to installing FastAI that works properly.