[PYTORCH] Hierarchical Attention Networks for Document Classification

Introduction

Here is my pytorch implementation of the model described in the paper Hierarchical Attention Networks for Document Classification paper.

An example of app demo for my model's output for Dbpedia dataset.

An example of my model's performance for Dbpedia dataset.

How to use my code

With my code, you can:

Train your model with any dataset
Given either my trained model or yours, you could evaluate any test dataset whose have the same set of classes
Run a simple web app for testing purpose

Requirements:

python 3.6
pytorch 0.4
tensorboard
tensorboardX (This library could be skipped if you do not use SummaryWriter)
numpy

Datasets:

Statistics of datasets I used for experiments. These datasets could be download from link

Dataset	Classes	Train samples	Test samples
AG’s News	4	120 000	7 600
Sogou News	5	450 000	60 000
DBPedia	14	560 000	70 000
Yelp Review Polarity	2	560 000	38 000
Yelp Review Full	5	650 000	50 000
Yahoo! Answers	10	1 400 000	60 000
Amazon Review Full	5	3 000 000	650 000
Amazon Review Polarity	2	3 600 000	400 000

Additionally, I also use word2vec pre-trained models, taken from GLOVE, which you could download from link. I run experiments with all 4 word2vec files (50d, 100d, 200d and 300d). You could easily switch to other common word2vec models, like the one provided in FastText link In the paper, it is said that a pre-trained word2vec is used. However, to the best of my knowledge, at least in pytorch, there is no implementation on github using it. In all HAN github repositories I have seen so far, a default embedding layer was used, without loading pre-trained word2vec model. I admit that we could still train HAN model without any pre-trained word2vec model. However, to serve the purpose of re-implementing origin model, in all experiments, as mentioned above, I used 1 out of 4 pre-trained word2vec models as initilization for embedding layer.

Setting:

During my experiments, I found out that given different datasets and different embedding layer's dimension, some combinations of batch size and learning rate yield better performance (faster convergence and higher accuracy) than others. Particularly in some cases, if you set wrong values for these 2 very important parameters, your model will never converge. Detail setting for each experiments will be shown in Experiments part. I have not set a fixed number of epoches for each experiment. Instead, I apply early stopping technique, to stop training phase after test loss has not been improved for n epoches.

Training

If you want to train a model with default parameters, you could run:

python train.py

If you want to train a model with your preference parameters, like optimizer and learning rate, you could run:

python train.py --batch_size batch_size --lr learning_rate: For example, python train.py --batch_size 512 --lr 0.01

If you want to train a model with your preference word2vec model, you could run:

python train.py --word2vec_path path/to/your/word2vec

Test

For testing a trained model with your test file, please run the following command:

python test.py --word2vec_path path/to/your/word2vec, with the word2vec file is the same as the one you use in training phase.

You could find some trained models I have trained in link

Experiments:

Results for test set are presented as follows: A(B/C):

A is accuracy.
B is learning rate used.
C is batch size.

Each experiment is run over 10 epochs.

GLOVE word2vec	50	100	200	300
ag_news	updated soon	updated soon	updated soon	updated soon
sogu_news	updated soon	updated soon	updated soon	updated soon
db_pedia	updated soon	updated soon	updated soon	updated soon
yelp_polarity	updated soon	updated soon	updated soon	updated soon
yelp_review	updated soon	updated soon	updated soon	updated soon
yahoo_answer	updated soon	updated soon	updated soon	updated soon
amazon_review	updated soon	updated soon	updated soon	updated soon
amazon_polarity	updated soon	updated soon	updated soon	updated soon

The training/test loss/accuracy curves for each dataset's experiments (with the order from left to right, top to bottom is 50d, 100d, 200d and 300d word2vec) are shown below:

ag_news

db_pedia

yelp_polarity

yelp_review

Yahoo! Answers

amazon_review

amazon_polarity

There are some experiments I have not had time to train. For such experiments, statistics as well as loss/accuracy visualization are empty. Additionally, there are some other experiments, I can not wait until they are finished, hence I stopped training phase before it should be . You could see whether a model was stopped by early stopping technique or by me by looking at the test loss curve, if the loss is not improved for 5 consecutive epoches, it is the former case. Othewise, if the loss is still going down, it is the latter case. When I have time, I will complete the incomplete experiments, and update results here.

After completing training phase, you could see model's parameters you have set, accuracy, loss and confusion matrix for test set at the end of each epoch at root_folder/trained_models/logs.txt. One example is shown below:

An example of logs.txt for Dbpedia dataset.