So I have a small doubt regarding saving embeddings. I would like to train a masked language model on some text corpus and then save the final trained embeddings weights to load into a NMT decoder. Is this possible? Otherwise how can I go about saving embeddings from a trained model? Thank you in advance!
Which version of OpenNMT are you using?
I just cloned the repository last week so it should be the latest version from GitHub (OpenNMT-py)
This is quite similar to:
There is no pre-existing way to do such a thing, but it shouldn’t be very difficult to add the few lines needed to dump those, or extract them from an existing checkpoint.
You can access all the parameters directly by loading a checkpoint:
checkpoint = torch.load("my_checkpoint.pt")
checkpoint is a python
dict containing various entries, one of which is
'model', which is the
state_dict of the pytorch model.