Sabotage data is inconsistent when training

[krogenth]

Currently, we convert all recorded data into dictionaries to avoid unnecessary load times performing conversions while defaulted data uses the data.proto_defaults definitions where applicable.

This becomes an issue when starting to convert during training, where a recorded sabotage is a dictionary goes through convert_dict, but the defaulted TaskData goes through convert_taskdata, resulting in misaligned array dimensions.

Attached is a recording with some sabotage data if necessary for testing.

133264897257692422.zip

[krogenth]

Looked into it more, a run without the decoded version of the gymbag2 file will run perfectly fine, but attempting to run with the decoded version already made will result in the issue described.

So we'll need to eventually normalize data between these two versions, where that should be, I don't know.

[Alexejhero]

Related(?): https://github.com/danielgtaylor/python-betterproto/issues/475

[krogenth]

Depends on what we want to have our data as. My understanding was that we wanted to be using the betterproto definitions made, but always using them was too costly to convert to. If we actually want to be using the attr_dict, we'd get unnecessary data fed into the model(mostly the id and type of the task).

Reading through the linked issue, yes, that does seem to be related.

[Alexejhero]

So the problem is that the two datasets are different, the easiest solution would be to load the dictionary even in the first run, when first caching the file.

This however brings a new problem, everything is just dictionaries instead of types and the convert functions lose meaning.

[krogenth]

We might also need to covert from proto_defaults to dict defaults as well, unless there's a non-performance heavy way to treat/convert the proto classes as dict.

Currently, if there isn't a definition, we use the default, which won't match.

[Alexejhero]

This is no longer a problem as we're now saving the entire Game as a pickle file.