load_archive and model from_archive

[vikigenius]

Looking at this example:

allennlp/test_fixtures/basic_classifier/experiment_from_archive.jsonnet

Lines 8 to 11 in f877fdc

	"model": {
	"type": "from_archive",
	"archive_file": "test_fixtures/basic_classifier/serialization/model.tar.gz",
	},

When I use this for training, it produces an archive where during inference time when I try to use to load_archive it needs the original archive to be there as well.

Can this be avoided?

I basically want to further train a model from a previous archive a1 using new data. So I use the above snippet.
But this produces an archive a2 where the config.json mentions a1. Which means during inference/prediction time I need both a1 and a2 archives in my directory structure. When technically I shouldn't be needing a1.

How do I avoid this dependency on a1 without manually modifying the a2 archive to change the model definition.

[epwalsh]

Hey @vikigenius, see #5604 (comment).

Long story short, the easiest way around this is to manually modify the config.json file within the model.tar.gz to remove any references to the old archive.