support marian

[shuoyangd]

shouldn't be hard :)

[mjpost]

This is a good idea.

One thing we also might consider is allowing all the decoder tape files to be loaded at the same time. This would require renaming tasks like "decode" to "sockeye_decode" and creating a variable for downstream constraints. The nice thing, though, is that you could branch on which decoder to use, and therefore easily build contrastive baselines.

Might not be worth the trouble, though...

[shuoyangd]

IMO that's a bit too much a corner-case scenario -- even if such scenario exists, one may as well clone 3 different copies of the pipeline and run 3 experiments with different decoders imported.

There's that decoder option mapping issue as well -- different decoders will have same options for some of the stuff, and different ones for others. Normalizing these options sounds like a daunting task to me.