Following the pointers from @romungi-MSFT, defining estimator with gpubase image; "mcr.microsoft.com/azureml/base-gpu:openmpi3.1.2-cuda10.1-cudnn7-ubuntu18.04" solves the problem, and Tensorflow 2.2 can be included. Tensorflow uses GPU by default when available.
estimator = Estimator(source_directory=experiment_folder,
compute_target=compute_target,
script_params=script_params,
entry_script='rps_efn_b0.py',
node_count=1,
conda_packages=['ipykernel'],
pip_packages = ['azureml-sdk',
'pyarrow',
'pyspark',
'azureml-mlflow',
'joblib',
'matplotlib',
'Pillow',
'tensorflow==2.2',
'tensorflow-datasets',
'tensorflow-hub',
'azureml-defaults',
'azureml-dataprep[fuse,pandas]'],
custom_docker_image='mcr.microsoft.com/azureml/base-gpu:openmpi3.1.2-cuda10.1-cudnn7-ubuntu18.04')