oliviersalaun

Hello, I didn't go further in the implementation of barthez. So far, I just added some lines for loading barthez tokenizer and checkpoint in `pretrain_nmt.py` though that was not enough...