What docker image should I pull? [train on a Ubuntu PC, inference on Jeston]

Accelerated Computing NGC GPU Cloud Docker and NVIDIA Docker
1

What docker image should I pull in my situation?
This is my vision:
Train a model on Ubuntu PC, and Inference on Jetson

  1. [Ubuntu PC] Train with original data using Tensorflow in docker
  2. [Ubuntu PC → Jetson nano] Export the Tensor RT model to Jetson nano.
  3. [Jetson nano] Convert the model to TensorRT format
    4.[Jetson nano] Inference image use with the Tensor RT model.

Device:
Ubuntu 18.04 + RTX3060 : for training
Jetson nano [Jetpack 4.6]: for inference

My question:

  1. What docker image should I pull on Ubuntu PC? Nvidia:tensorflow series is good for it?
    https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow
    https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html#framework-matrix-2020

  2. What is key point to select?
    Consider with [TensorFlow for Jetson Compatibility] or [Cuda version] on Jetson nano?
    If Tensorflow compatibility is key, I should select 21.07~21.09 but its cuda is ver 11.
    If Cuda version is key, I should select 19.xx but Jetpack 4.6 is not compatible to TensorFlow for Jetson.