search

facebookresearch / SimulEval

SimulEval: A General Evaluation Toolkit for Simultaneous Translation
Creative Commons Attribution Share Alike 4.0 International
102 stars 36 forks source link

ReadAction and WriteAction #79

Closed ZechengLi19 closed 1 year ago

ZechengLi19 commented 1 year ago

Thanks for your awesome work. However, I am confused by how ReadAction and WriteAction affect model input and output. I can't find this part of the code.

Thanks!

xutaima commented 1 year ago

Hi @ZechengLi19, sorry for the late response.

When a ReadAction is returned, the model will not generate output. Instead, more source input will be read and the states (e.g. states.source will include the next input) will be updated before the next time the policy function is called.

When a WriteAction is returned, the model will generate output. The evaluator will receive the output and track the delay of the output. The output will also update the states (e.g. states.target)

Hope this answers your question.

ZechengLi19 commented 1 year ago

Hi @ZechengLi19, sorry for the late response.

When a ReadAction is returned, the model will not generate output. Instead, more source input will be read and the states (e.g. states.source will include the next input) will be updated before the next time the policy function is called.

When a WriteAction is returned, the model will generate output. The evaluator will receive the output and track the delay of the output. The output will also update the states (e.g. states.target)

Hope this answers your question.

This simple and clear answer helped me a lot!

Thanks for your reply!