No batch-training? about editsql HOT 4 CLOSED

Hi, Longxu,

Thanks! Our implementation uses only 1 interaction in the interaction_batch. This is because (1) it is hard to implement batch_size>1 for interactions, and (2) memory could be an issue when batch_size>1.

The batch_size is used to normalize the loss of each example by dividing the batch_size (https://github.com/ryanzhumich/editsql/blob/master/model/schema_interaction_model.py#L466)

Implementing batch_size>1 for Spider (which is not interaction) is possible, but not available yet...

Best,
Rui

from editsql.

Hi, Rui

Thanks for your prompt reply!
And is there an independent model for Spider? Currently, all those three datasets are trained with the model scheme_interaction_model. The function train_epoch_with_utterances trains model for a single epoch given batches of utterance data, but no model has been called in run.py.

Best,
Longxu

from editsql.

You are welcome. You are correct, train_epoch_with_utterances could be used for Spider. But sorry we currently do not have such an implementation.

from editsql.

Thanks for your job!

from editsql.