Question about Multi-task Configuration

[feralvam]

Hi,
I just read your paper and it's a very nice work. I have a 'practical' question that I hope you can help me with. Where in the released code is the multi-task configuration? I couldn't find where you tell sockeye (using the command-line) to sum up the losses of each task before backprop. I'm not familiar with sockeye.

Thank you.

[sweta20]

Hi,

The way we do this is, we don't actually sum up the individual loss components as these are calculated on different datasets but share the same seq2seq model.

So, for example to train the multitask system illustrated in the diagram, you can use
bash scripts/main.sh -d -b -n -e -f -g -i <iter>
You can check the description of the arguments using.
bash scripts/main.sh -h

I hope this makes it slightly clear.