Update TensorFlow version to 2.2

Hi there!!

I’m gathering some feebacks from the community in order to know what you guys think to fix the minimum TensorFlow version to 2.2?

This would allow the models to be compliant with several cool features such as:

  • Model.train_step method to create simple custom training loop that can be used with Model.fit
  • Pre-computed SavedModel inputs/outputs with Model._saved_model_inputs_spec
  • Make use of the new tf.data.experimental.assert_cardinality that can fix a size to a TFRecord or from_generator TensorFlow Datasets.
  • Drop the Python 2 support such as transformers itself.

Let me know here all your suggestions about this :slight_smile: