Skip to content

Latest commit

 

History

History
20 lines (16 loc) · 1.11 KB

new_task_type.md

File metadata and controls

20 lines (16 loc) · 1.11 KB
Raw

Adding a new task-type

back to main README

Adding a new task type can involve some effort, so I suggest to get in touch with robv@itu.dk.

Basically it involves four steps:

  • Check if an existing dataset reader can match your task type; if it does, make sure it uses the correct one: read_function in machamp/data/machamp_dataset.py. If there is no reader that matches your data structure, you have write a new one.
  • Add the task-type to the configs/param.json, including any parameters you need to use for the following step
  • The main task is to include a task decoder. If you are lucky, you can draw inspiration from the AllenNLP models. I would suggest to take a look at the most similar of the existing decoders in machamp/models/.
  • Make sure the MachampModel model forward pass passes the right encoding (sentence/word-level) to the right decoder.
  • Add conversion of gold labels to batches in machamp/utils/myutils.py in prep_batch
  • Prediction might not work by default, and might need some adaptation in machamp/predictor/predict.py