From @daikeshi #127 (comment):
[tfr-read] assumes that there'll be _schema.pb or _inferred_schema.pb file stored with the dataset, but I think we are now moving to use stats file or schema txt, we should probably change tfr-read to reflect that.
tfr-read should output json data:
- using schema txt file if it's provided
- using stats file if there's a stats file with the data, but no explicit schema txt file argument
- using the default ExampleDecoder if no schema file and stats file are provided