README.md

Hierarchical Attention Networks

Abstract

An implementation of Hierarchical Attention Networks for Document Classification in PyTorch.

Installation

You can install the package from pip:

pip install hierarchical-attention-networks

Requirements

You can see the requirements in setup.cfg.

Usage

This package provides two neural networks. The first one is SentenceModel in han.model.sentence. It implements a word encoder and a word attention. The second one is DocumentModel in han.model.document. It depends on SentenceModel, and implements a sentence attention and a sentence encoder.

SentenceModel.forward takes a list of torch.Tensors. A tensor represents to a sentence and is the index of the words in the sentence. It returns a tuple of two tensors. The first one is the sentence embeddings, and its shape is (num of sentences, self.sentence_dim). The second one represents the attention. The shape is (the length of the longest tensor in an input, num of sentences).

DocumentModel.forward takes documents. A document is a list of tensors, and a tensor represents a sentence. It returns a quadruple. the first item represents document embeddings. The second and third items represend sentence attention and word attention. The fourth items is a list of the numbers of the sentences in a document.

You can instantiate them by SentenceModelFactory DocumentModelFactory. They can accept pretrained word embeddings.

Example

You can fit a model that depends on DocumentModel on AG News by the following comands.

import han.example.document as d
import torchtext.vocab as v
import torch
d.train(
    "d_enc.pth",
    "d_model.pth",
    device=torch.device("cpu"),
    embedding_sparse=False, 
	pre_trained=v.FastText(),
)

Test

You can run tests from the command line.

pip install -e .[dev]
python -m unittest