If DeepSpeech migrated to tf r2.2 in release 0.8.0 then why is the training docker container for 0.9.3 still using tf r1.15? I am trying to use tensorflow addons but averaging model checkpoints but those don’t work with tf r1.15 and if I try to upgrade the tf version inside the docker then the code breaks.
You are mixing training and inference. Only inference is running r2.3 as of 0.9 / master.
Upgrading training is still an open problem if you want to help: https://github.com/mozilla/DeepSpeech/pull/3485