Multidataset model training

[ElliotStein]

Hi all,

Firstly, congrats on the excellent research!

I would like to reproduce Table 3 of the paper and need to first train the models to be merged. However, mmultids_trainer.py requires a checkpoint_path argument which doesn't make sense to me. I thought the idea was to train each model from scratch, not a pre-trained or shared initialisation? Why do we need this argument?

As a side note, if possible, it would be great if you could make pre-trained weights publicly available to make reproducibility more efficient.

Thanks!