Retrieve NMT Encoder's output

[brash6]

Hi,

Assuming that I've trained a NMT with OpenNMT, how coud I retrieve the representation vector of new sequences ?

More precisely, If I trained a NMT to translate from French to English, if I want to predict the output sequence for a new French sentence, how could I also retrieve the output vector of the encoder ?

Thank you !

[francoishernandez]

Hi there, This is quite similar to this post. It is not handled "as is" but it should not be too difficult to add it if you're willing to dive a bit in the code.

[brash6]

Thank you for your answer, I'll post here If I manage to do it !

[Marvinmw]

hi, do you manage to do it? @brash6

[brash6]

Hello, not really, If you have any idea how to do it, I am interested !

[shashankg7]

Hi,

Did anybody had any success in getting the encoder representation from the trained model?

[francoishernandez]

This should not be very difficult guys, if you're willing to dive a bit in the code. You can start by printing / torch.save what's happening here, after the encoder forward: https://github.com/OpenNMT/OpenNMT-py/blob/8b073fb2a047509ff590839b1194a155ec1a50bf/onmt/models/model.py#L45

Then, it's just plumbing to retrieve this at the translator level.

[shashankg7]

Thanks for your reply.

I have mainly used it's command line tool only. Have not used the torch api. Any reference to building a transformer translation model with its pytorch api?

Pardon if I am missing anything obvious here.

On Thu, Oct 8, 2020, 1:59 PM François Hernandez [email protected] wrote:

This should not be very difficult guys, if you're willing to dive a bit in the code. You can start by printing / torch.save what's happening here, after the encoder forward:

https://github.com/OpenNMT/OpenNMT-py/blob/8b073fb2a047509ff590839b1194a155ec1a50bf/onmt/models/model.py#L45

Then, it's just plumbing to retrieve this at the translator level.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/OpenNMT/OpenNMT-py/issues/1811#issuecomment-705416630, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA5VTOB6KRTPNAMYAH5MCNLSJVZ7TANCNFSM4NZGBKTQ .

[francoishernandez]

You don't need to rebuild everything just to retrieve some representations. You can just edit the code and run the scripts.

[shashankg7]

Thank you for your prompt response.

I have one more doubt. Please pardon me if my doubts appear naive.

The installed scripts will invoke the code from the installed location, right?

How do I ensure that the scripts call the changes I make in the onmt codebases' model?

[francoishernandez]

Easiest is probably to git clonethe repo in the location you'd like and work from there.

[Shikhar-S]

Hi @francoishernandez! Is anyone working on this issue? I want to contribute.