Inverse relation models only works with ConvE about kge HOT 7 CLOSED

You might have to change the lookup_embedder.initialize to normal_ to see the relevant error. Also, specifying the types for inverse_relations_model.base_model.entity_embedder and relation_embedder fix this, but ideally we shouldn't have to.

from kge.

I do not think this is related to the commit you mention. Instead, it seems to be an error in your config file: lookup_embedder.initialize: uniform is set in the config, but lookup_embedder.initialize_args is not set. It therefore complains. Adding lookup_embedder.initialize_args.a=1. (for example) resolves the problem.

from kge.

from kge.

No, the error message was good: it complained that the a argument of the initializer was missing. (Please include the error messages in the issue next time.)

from kge.

Ok, so two things happened here. 1) The config file I uploaded was not correct, sorry. I attach a new one. 2) The issue was hackily fixed in 7066d9b, because I needed to run searches already. So, in order to reproduce, one needs to go back to 38a6968 and run the attached config file. The error message is:

KeyError: 'entity_embedder'

Apparently, we did not notice because the entity_embedder and relation_embedder entries are already added for ConvE when adding the hack for the entity_bias.

toy-complex-inverse-train.yaml.zip

from kge.

The problem was not in the configuration, but in the implementation of KgeEmbedder. Fixed & hack removed. Please reopen if there are still problems.

from kge.

BTW: We should post full stack traces in the future; the trace here showed where the error occurred.

from kge.