A framework agnostic Python NLP library for data loading and preprocessing.
Podium is a framework agnostic Python natural language processing library which standardizes data loading and preprocessing. Our goal is to accelerate users' development of NLP models whichever aspect of the library they decide to use.
We desire Podium to be lightweight, in terms of code and dependencies, flexible, to cover most common use-cases and easily adapt to more specific ones and clearly defined, so new users can quickly understand the sequence of operations and how to inject their custom functionality.
Check out our documentation for more details. The main source of inspiration for Podium is an old version of torchtext.
You can install podium using pip
pip install podium-nlp
Commands to install podium from source
git clone [email protected]:TakeLab/podium.git && cd podium pip install .
For more detailed installation instructions, check the installation page in the documentation.
Use some of our pre-defined datasets:
>>> from podium.datasets import SST >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits() >>> sst_train.finalize_fields() # Trigger vocab construction >>> print(sst_train) SST({ size: 6920, fields: [ Field({ name: text, keep_raw: False, is_target: False, vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 16284}) }), LabelField({ name: label, keep_raw: False, is_target: True, vocab: Vocab({specials: (), eager: False, is_finalized: True, size: 2}) }) ] }) >>> print(sst_train[222]) # A short example Example({ text: (None, ['A', 'slick', ',', 'engrossing', 'melodrama', '.']), label: (None, 'positive') })
Load datasets from 🤗 datasets:
>>> from podium.datasets.hf import HFDatasetConverter as HF >>> import datasets >>> # Load the huggingface dataset >>> imdb = datasets.load_dataset('imdb') >>> print(imdb.keys()) dict_keys(['train', 'test', 'unsupervised']) >>> # Wrap it so it can be used in Podium (without being loaded in memory!) >>> imdb_train, imdb_test, imdb_unsupervised = HF.from_dataset_dict(imdb).values() >>> # We need to trigger Vocab construction >>> imdb_train.finalize_fields() >>> print(imdb_train) HFDatasetConverter({ dataset_name: imdb, size: 25000, fields: [ Field({ name: 'text', keep_raw: False, is_target: False, vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 280619}) }), LabelField({ name: 'label', keep_raw: False, is_target: True }) ] })
Load your own dataset from a standardized tabular format (e.g. csv, tsv, jsonl, ...):
>>> from podium.datasets import TabularDataset >>> from podium import Vocab, Field, LabelField >>> fields = {'premise': Field('premise', numericalizer=Vocab()), ... 'hypothesis':Field('hypothesis', numericalizer=Vocab()), ... 'label': LabelField('label')} >>> dataset = TabularDataset('my_dataset.csv', format='csv', fields=fields) >>> dataset.finalize_fields() # Trigger vocab construction >>> print(dataset) TabularDataset({ size: 1, fields: [ Field({ name: 'premise', keep_raw: False, is_target: False, vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 15}) }), Field({ name: 'hypothesis', keep_raw: False, is_target: False, vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 6}) }), LabelField({ name: 'label', keep_raw: False, is_target: True, vocab: Vocab({specials: (), eager: False, is_finalized: True, size: 1}) }) ] })
Check our documentation to see how you can load a dataset from Pandas, the CoNLL format, or define your own Dataset subclass (tutorial coming soon).
We wrap dataset pre-processing in customizable Field classes. Each Field has an optional Vocab instance which automatically handles token-to-index conversion.
>>> from podium import Vocab, Field, LabelField >>> vocab = Vocab(max_size=5000, min_freq=2) >>> text = Field(name='text', numericalizer=vocab) >>> label = LabelField(name='label') >>> fields = {'text': text, 'label': label} >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits(fields=fields) >>> sst_train.finalize_fields() >>> print(vocab) Vocab({specials: ('<UNK>', '<PAD>'), eager: False, finalized: True, size: 5000})
Each Field allows the user full flexibility to modify the data in multiple stages:
- Prior to tokenization (by using pre-tokenization
hooks) - During tokenization (by using your own
tokenizer) - Post tokenization (by using post-tokenization
hooks)
You can also completely disregard our preprocessing and define your own by setting your own numericalizer.
You could decide to lowercase all the characters and filter out all non-alphanumeric tokens:
>>> def lowercase(raw): ... return raw.lower() >>> def filter_alnum(raw, tokenized): ... filtered_tokens = [token for token in tokenized if ... any([char.isalnum() for char in token])] ... return raw, filtered_tokens >>> text.add_pretokenize_hook(lowercase) >>> text.add_posttokenize_hook(filter_alnum) >>> fields = {'text': text, 'label': label} >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits(fields=fields) >>> sst_train.finalize_fields() >>> print(sst_train[222]) Example({ text: (None, ['a', 'slick', 'engrossing', 'melodrama']), label: (None, 'positive') })
Pre-tokenization hooks accept and modify only on raw data.
Post-tokenization hooks accept and modify raw and tokenized data.
A common use-case is to incorporate existing components of pretrained language models, such as BERT. This is extremely simple to incorporate as part of our Fields. This snippet requires installation of the 🤗 transformers (pip install transformers) library.
>>> from transformers import BertTokenizer >>> # Load the tokenizer and fetch pad index >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased') >>> pad_index = tokenizer.convert_tokens_to_ids(tokenizer.pad_token) >>> # Define a BERT subword Field >>> subword_field = Field(name="subword", ... padding_token=pad_index, ... tokenizer=tokenizer.tokenize, ... numericalizer=tokenizer.convert_tokens_to_ids) >>> fields = {'text': subword_field, 'label': label} >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits(fields=fields) >>> # No need to finalize since we're not using a vocab! >>> print(sst_train[222]) Example({ subword: (None, ['a', 'slick', ',', 'eng', '##ross', '##ing', 'mel', '##od', '##rama', '.']), label: (None, 'positive') })
For a more interactive introduction, check out the quickstart on Google Colab:
Full usage examples can be found in our docs under the Examples heading.
We welcome contributions! To learn more about making a contribution to Podium, please see our Contribution page and our Roadmap.
We use SemVer for versioning. For the versions available, see the tags on this repository.
- Podium is currently maintained by Ivan Smoković, Mario Šaško, Filip Boltužić, and Martin Tutek. A non-exhaustive but growing list of collaborators: Silvije Skudar, Domagoj Pluščec, Marin Kačan, Dunja Vesinger, Mate Mijolović.
- Thanks to the amazing Mihaela Bošnjak for the logo!
- Project made as part of TakeLab at Faculty of Electrical Engineering and Computing, University of Zagreb.
See also the list of contributors who participated in this project.
If you are using Podium, please cite the following entry in your work:
@misc{tutek-etal-2021-podium,
author = {Martin Tutek and Filip Boltužić and Ivan Smoković and Mario Šaško and Silvije Škudar and Domagoj Pluščec and Marin Kačan and Dunja Vesinger and Mate Mijolović and Jan Šnajder},
title = {Podium: a framework-agnostic NLP preprocessing toolkit},
year = {2021},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/TakeLab/podium}},
commit = {4fed78b8d8366768df10454b8368f416a3305cc4}
}
This project is licensed under the BSD 3-Clause - see the LICENSE file for details.