Reload Pipeline Results From Directory

[mberr]

Problem Statement

Reloading the pykeen.pipeline.PipelineResult object from a directory seems to be frequently desired, cf., e.g.,

• How to evaluate model results from a pre-trained model? #1181 (reply in thread)
• Load pipeline result from save_to_directory #1189

Describe the solution you'd like

Add a classmethod, e.g., load_from_directory to pykeen.pipeline.PipelineResult which allows to restore the object from the information stored in the directory.

Describe alternatives you've considered

An alternative could be to improve the documentation on how to load the relevant parts into, e.g., pandas Dataframes.

Additional information

• The serializaton format is described here: https://pykeen.readthedocs.io/en/latest/api/pykeen.pipeline.PipelineResult.html#pykeen.pipeline.PipelineResult.save_to_directory

Issue Template Checks

• This is not a bug report (use a different issue template if it is)
• This is not a question (use the discussions forum instead)

[nicolafan]

Hello! I was looking into this, and I already have an idea about how to recover the PipelineResult from the files saved to the directory.

How should we treat the following attributes, whose values are not saved to the results directory?

• random_seed
• training_loop
• configuration
• version
• git_hash

Everything else can be recovered (I've just a doubt about how to recover the stopper, while I've already seen how to do that for the other attributes). Thank you for your help!

[mberr]

Can we make them optional (i.e., allow providing None for those in case they are not present)? This would allow to load results which have been saved before the current change.

For future save_to_directory calls, we could store them, too.