When continuing a training, the option for embeddeing seems to have no effect

Support
1

When training an existing model and using the options: -pre_word_vecs_enc, -pre_word_vecs_enc, putting a fake path for the embedding files, does not throw ‘no such file or directory’ error, the training continues.

When it’s a new training, the error is thrown when it’s a fake path.

onmt_train -data /data/demo20k -save_model enru20k-model
-train_from /data/enru20k-model_step_100000.pt
-pre_word_vecs_enc /fake_path/embeddings.enc.pt
-pre_word_vecs_dec /fake_path/embeddings.dec.pt
-train_steps 200000

2

True, see the code here.

Not sure it would make sense to change the embeddings along with a pretrained checkpoint. Your pretrained weights would not necessarily be adapted to the new embeddings. I’ve not done many experiments with either pretraining or pretrained embeddings so I may be wrong here.

Anyways, you can probably easily adapt this code to test the behavior and tell us if this has any interesting result!