grouphug

GroupHug is a library with extensions to 🤗 transformers for multitask language modelling. In addition, it contains utilities that ease data preparation, training, and inference.

Overview

The package is optimized for training a single language model to make quick and robust predictions for a wide variety of related tasks at once, as well as to investigate the regularizing effect of training a language modelling task at the same time.

You can train on multiple datasets, with each dataset containing an arbitrary subset of your tasks. Supported tasks include:

• A single language modelling task (Masked language modelling, Masked token detection, Causal language modelling).
  • The default collator included handles most preprocessing for these heads automatically.
• Any number of classification tasks, including single- and multi-label classification and regression
  • A utility function that automatically creates a classification head from your data.
  • Additional options such as hidden layer size, additional input variables, and class weights.
• You can also define your own model heads.

Quick Start

The project is based on Python 3.8+ and PyTorch 1.10+. To install it, simply use:

pip install grouphug

Documentation

Documentation can be generated from docstrings using make html in the docs directory, but this is not yet on a hosted site.

Example usage

import pandas as pd
from datasets import load_dataset
from transformers import AutoTokenizer

from grouphug import AutoMultiTaskModel, ClassificationHeadConfig, DatasetFormatter, LMHeadConfig, MultiTaskTrainer

# load some data. 'label' gets renamed in huggingface, so is better avoided as a feature name.
task_one = load_dataset("tweet_eval",'emoji').rename_column("label", "tweet_label")
both_tasks = pd.DataFrame({"text": ["yay :)", "booo!"], "sentiment": ["pos", "neg"], "tweet_label": [0,14]})

# create a tokenizer
base_model = "prajjwal1/bert-tiny"
tokenizer = AutoTokenizer.from_pretrained(base_model)

# preprocess your data: tokenization, preparing class variables
formatter = DatasetFormatter().tokenize().encode("sentiment")
# data converted to a DatasetCollection: essentially a dict of DatasetDict
data = formatter.apply({"one": task_one, "both": both_tasks}, tokenizer=tokenizer, test_size=0.05)

# define which model heads you would like
head_configs = [
    LMHeadConfig(weight=0.1),  # default is BERT-style masked language modelling
    ClassificationHeadConfig.from_data(data, "sentiment"),  # detects dimensions and type
    ClassificationHeadConfig.from_data(data, "tweet_label"),  # detects dimensions and type
]
# create the model, optionally saving the tokenizer and formatter along with it
model = AutoMultiTaskModel.from_pretrained(base_model, head_configs, formatter=formatter, tokenizer=tokenizer)
# create the trainer
trainer = MultiTaskTrainer(
    model=model,
    tokenizer=tokenizer,
    train_data=data[:, "train"],
    eval_data=data[["one"], "test"],
    eval_heads={"one": ["tweet_label"]},  # limit evaluation to one classification task
)
trainer.train()

Tutorials

See examples for a few notebooks that demonstrate the key features.

Supported Models

The package has support for the following base models:

• Bert, DistilBert, Roberta/DistilRoberta, XLM-Roberta
• Deberta/DebertaV2
• Electra
• GPT2, GPT-J, GPT-NeoX, OPT

Extending it to support other models is possible by simply inheriting from _BaseMultiTaskModel, although language modelling head weights may not always load.

Limitations

• The package only supports PyTorch, and will not work with other frameworks. There are no plans to change this.
• Grouphug was developed and tested with 🤗 transformers 4.19-4.22. We will aim to test and keep compatibility with the latest version, but it is still recommended to lock the latest working versions.

See the contributing page if you are interested in contributing.

License

grouphug was initially developed at Chatdesk and is licensed under the Apache 2 license.