I am experimenting with eole training for bilingual MT models. I can successfully train a model with eole from scratch using wmt17 recipe. However I also want to see if I can fine-tune my legacy opennmt-py models with eole. With opennmt, it was possible to load a model checkpoint using train_from parameters. But for eole it seems from code that it expects a model directory with checkpoint, config.json and vocab.json files. Is there a way to convert opennmt checkpoint files (.pt) to this required format?
Hi,
I am experimenting with eole training for bilingual MT models. I can successfully train a model with eole from scratch using wmt17 recipe. However I also want to see if I can fine-tune my legacy opennmt-py models with eole. With opennmt, it was possible to load a model checkpoint using
train_from
parameters. But for eole it seems from code that it expects a model directory with checkpoint, config.json and vocab.json files. Is there a way to convert opennmt checkpoint files (.pt) to this required format?Thanks