The easiest way to get Marian-trained models into ELG is as follows:
Create a Docker image with an ELG-compatible REST-based translation server that incorporates the respective model. This is easy. See https://github.com/ugermann/marian-docker for code and details. In a nutshell:
Put all relevant files (vocab.spm, model.bin) into a separate model directory.
Create a decoder.yml file (with a few extra fields) for your setup in the model directory.
Copy the appropriate Dockerfile into the model directory.
Run docker build -t /path/to/your/model/directory
Push the image to a Docker repository of your choice
Announce the resource to ELG. For this, you'll need to provide a metadata record to ELG. This process is currently still a bit rough around the edges, but ILSP are working hard to make it much easier. Currently, you'll need to provide the metadata as an XML file that confirms to a specific DTD. Penny is the person to talk to about creating and ingesting resource metadata records into ELG.
develop OPUS-MT docker for ELG (based on https://github.com/ugermann/marian-docker?). Information from Ullrich Germann:
The easiest way to get Marian-trained models into ELG is as follows:
Create a Docker image with an ELG-compatible REST-based translation server that incorporates the respective model. This is easy. See https://github.com/ugermann/marian-docker for code and details. In a nutshell: Put all relevant files (vocab.spm, model.bin) into a separate model directory. Create a decoder.yml file (with a few extra fields) for your setup in the model directory. Copy the appropriate Dockerfile into the model directory. Run docker build -t /path/to/your/model/directory
Push the image to a Docker repository of your choice
Announce the resource to ELG. For this, you'll need to provide a metadata record to ELG. This process is currently still a bit rough around the edges, but ILSP are working hard to make it much easier. Currently, you'll need to provide the metadata as an XML file that confirms to a specific DTD. Penny is the person to talk to about creating and ingesting resource metadata records into ELG.