Skip to content

Latest commit

 

History

History
65 lines (52 loc) · 2.67 KB

ROADMAP.md

File metadata and controls

65 lines (52 loc) · 2.67 KB
Raw

Roadmap

If you are interested in making a contribution to Podium, this page outlines some changes we are planning to focus on in the near future. Feel free to propose improvements and moficiations either via discussions or by raising an issue.

Order does not reflect importance.

Major changes

  • Dynamic application of Fields

    • Right now, for every change in Fields the dataset needs to be reloaded. The goal of this change would be to allow users to replace or update a Field in a Dataset. The Dataset should be aware of this change (e.g. by keeping a hash of the Field object) and if it happens, recompute all the necessary data for that Field.

    The current pattern is:

      # Load a dataset
  fields = {'text':text, 'label':label}
  dataset = load_dataset(fields=fields)

  # Decide to change something with one of the Fields
  text = Field(..., tokenizer=some_different_tokenizer)
  # Potentially expensive dataset loading is required again
  dataset = load_dataset(fields=fields)

    Dataset instances should instead detect changes in a Field and recompute values (Vocabs) for the ones that changed.

  • Parallelization

    • For data preprocessing (apply Fields in parallel)
    • For data loading

  • Conditional processing in Fields

    • Handle cases where the values computed in one Field are dependent on values computed in another Field

  • Experimental pipeline

    • podium.experimental, wrappers for model framework agnostic training & serving
    • Low priority

Minor changes

  • Populate hooks & preprocessing utilities
    • Lowercase, truncate, extract POS, ...
  • Populate pretrained vectors
    • Word2vec
    • Interface with e.g. gensim
  • Improve Dataset coverage
    • Data wrappers / abstract loaders for other source libraries and input formats
  • BucketIterator modifications
    • Simplify setting the sort key (e.g., in the basic case where the batch should be sorted according to the length of a single Field, accept a Field name as the argument)
  • Improve HF/datasets integration
    • Better automatic Field inference from features
    • Cover additional feature datatypes (e.g., image data)
    • Cleaner API?
  • Centralized and intuitive download script
    • Low priority as most data loading is delegated to hf/datasets
  • Add a Mask token for MLM (can be handled with posttokenization hooks right now, but not ideal)

Documentation

  • Examples
    • Language modeling
    • Tensorflow model
    • Various task types
  • Chapters
    • Handling datasets with missing tokens
    • Loading data from pandas / porting data to pandas
    • Loading CoNLL datasets
    • Implementing your own dataset subclass