How do we programmatically use this framework to train a semantic parser?

[entslscheia]

As a newbie to allennlp, my understanding of the paradigm of training a model using allennlp is that we should have a DatasetReader object to read the data from disk, an Iterator object to batch the training data (i.e., objects of Instance), and a Model object to set up the computational graphs. Then basically, we pass those objects and an optimizer (or some other options) to a Trainer object trainer = Trainer(...), and train the model using trainer.train().

However, for the semantic parsing framework, I didn't see any code related to allennlp.training.Trainer, and the behavior of DecoderTrainer seems totally different from Trainer in allennlp. Instead of taking a Model object as a parameter to train, a DecoderTrainer object is served as a parameter of a Model object (e.g., MaximumMarginalLikelihood is inside WikiTablesMmlSemanticParser) to compute the loss.

So how do we exactly train a semantic parser programmatically? Are we still using a allennlp.training.Trainer object to incorporate everything (i.e., the model, dataset and an iterator) and train the model by calling train()?

I have read part of the source codes and have a basic understanding of how the semantic parsing framework looks like but didn't see any concrete example that includes a complete pipeline to train a semantic parser. I don't know maybe my question is kind of naive or trivial, but still confused for now.

[matt-gardner]

We still use the Trainer. It sounds like you're writing your own script to train your model; I'd recommend using the allennlp train command instead. You can follow this tutorial to learn more about how that whole setup works. In this repo, you'll find DatasetReaders and Models, plus a bunch of code that gets used inside those for semantic-parsing-specific stuff, and that's it. The Trainer code lives in the main repo, and we train using allennlp train and a configuration file.