NVidia GPU cards with architectures later than Turing do not properly infer image classification when using transpiled/cached models. This requires updating of Tensorflow to 2.7 to resolve.

[Crichen]

Current release of ML.Net is fixed to Tensorflow 2.3.1 with CUDA and Cudnn support at 10. and 7.6 respectively. Models trained with this set up work natively with Turing based cards. Newer GPU cards transpile and cache the trained model to run on their specific architectures, Ampere, Ada, Hopper, etc. Although the trained models will load with newer architectures the outputs give fixed results irrespective of the inputs.

To be compatible with the newer Ampere based architecture the Tensorflow.Net dependency needs to be raised to 0.70.2, using Tensorflow 2.7. This allows support of CUDA 11.2 and Cudnn 8.1 and for the trained models to run natively on Ampere without the need for transpiling.

These updates are tested and ready to clean up and commit for code review.