Ex 2: model files and forward signature

[waffelbaellchen]

Hi, in the provided model code files (e.g. model_knrm.py), the signature of the forward method states that query and document should be of type Dict[str. torch.Tensor]. To my understanding, we expect query to be a dictionary with string-keys and torch.Tensor-values. However, in the forward method, the provided code accesses the values of query and document via
query["tokens"]["tokens"]
and
document["tokens"]["tokens"]

Maybe I misunderstand something, but shouldn't we access values via query["tokens"] or document["tokens"]? I do not quite understand how a torch.Tensor has a ["tokens"] entry (which we try to access via query["tokens"]["tokens"] ).

[sophiaalthammer]

Hi, I understand you confusion! The structure of query['tokens']['tokens'] is due how the IrTripleDatasetReader and the IrTupleDatasetReader (in data_loading.py) read in the textfields for the query_tokens, doc_pos_tokens and doc_neg_tokens. For each of the query, doc_pos, doc_neg it reads in a TextField (https://docs.allennlp.org/main/api/data/fields/text_field/) which is defined as TextFieldTensors = Dict[str, Dict[str, torch.Tensor]].

The given structure is correct and works with the predefined dataloaders, so in order to complete the missing parts, I would advise you to load a small example of data and look at the structure how it is read in.

Does that answer your question?

Best,
Sophia

[waffelbaellchen]

Yes, now it became clear, thank you!