Modernmt: Training from an existing model
Is there a way to launch a training on a new set of data, using an already trained model as starting point ?
It could be very useful, to specialize a generic model, or simply enrich the process explained in #426.
EtienneAb3d
All 4 comments
Not at the moment. However our training process is calls the tensor2tensor t2t-train command under the hood. So if there is a way for doing so in t2t, we can easily replicate this in our tool.
davidecaroselli
on 24 Dec 2018
I came across this one. It could be useful https://github.com/tensorflow/tensor2tensor/issues/588
mzeidhassan
on 26 Dec 2018
Hello,
in master branch I have just merged a branch from @nicolabertoldi that implements this feature. You can now use mmt create and specify an initial model (instead of random initialized net) with the option --from-model.
It is now possible to specialize a pre-trained generic model just like @EtienneAb3d 's use case!
davidecaroselli
on 27 Feb 2019
Great news opening a lot of possibilities !
:+1:
EtienneAb3d
on 27 Feb 2019
Related issues
LoekvanKooten
路
4Comments
LoekvanKooten
路
3Comments
rptl
路
6Comments
nonsequitir
路
3Comments
mzeidhassan
路
3Comments
Most helpful comment
Hello,
in
masterbranch I have just merged a branch from @nicolabertoldi that implements this feature. You can now usemmt createand specify an initial model (instead of random initialized net) with the option--from-model.It is now possible to specialize a pre-trained generic model just like @EtienneAb3d 's use case!