Repository for the project in the course 732A92 Text Mining at Linköping University 2019.
This is a project in the above mentioned course. This repository includes the source code (src), some notebooks for mingling around (notebooks) and the report itself (report).
The two classifier used in this project are a normal feed-forward network and an unidirecrional LSTM (due to performance reasons just unidirecrional). Both are trained using Word2Vec embeddings (without context) and Bert embeddings (output of the last layer, with context). Word2Vec is being trained while the Bert model is frozen due to it's size. Also DistilBERT has been used to improve the performance, as the original BERT model is quite large.
First clone the repository locally:
git clone git@github.com:flennic/text-mining-project.git
Next you need the data set itself as well as the word embeddings.
The dataset Amazon Reviews for Sentiment Analysis can be found on Google Drive and is originally provided by Xiang Zhang. The data is expected to be in a folder data/original/ but you can change the path in the settings.json if desired.
You will also need the Word2Vec embeddings, which can be found e.g. here. The embeddings are expected to be in data/embeddings/, which can also be changed in the settings.json file.
If you want your sweet automation, just do chmod +x fetch-data.sh and then ./fetch-data.sh and the script will download the data for you. Be aware that the script does not have any error handling, so use it with caution. It also assumes default paths.
For starting and training a model, you call main.py. The default settings fault in main.py will be overwritten by your settings file. If you don't specify an entry, the default from main.py is taken automatically.
You will find an example settings.json further down this readme.
I strongly suggest having at least 20GB of free disk space, 16GB of memory and a dedicated graphics card, otherwise you won't have much fun running this project. The default settings assume above mentioned requirements and 8GB of VRAM, if you have less, feel free to reduce batch sizes. Be aware that training times might drastically increase.
The preprocessing will take some time, so be patient. Leave the cache parameter set to true, so that you only have to do it once.
You should install CUDA to make the GPU work.
For the following I suggest setting up a new anaconda environment. You will need the following libraries:
- PyTorch 1.3.1 (probably also works with others, but you're probably best with this one for this project).
conda install -c pytorch pytorch
- The transformers library from huggingface.
pip install transformers
- gensim is used for managing Word2Vec embeddings. This is just for my personal convenience.
pip install gensim
Please be aware that this is just a small project aimed at two weeks of full time work, so the code might be edgy in some parts. Nor is this code intended to be used in a productive environment in any sense.
The printed and logged training accuracies are trailing accuracies, so you will see a small jump in accuracy when starting a new epoch, especially after the first one. This is due to the fact that in the beginning of each epoch the accuracy is lower compared to towards the end. As the training set is rather large, you will see this behaviour.
For further details, please check the paper.
{
"orig_train_path": "data/original/train.csv",
"orig_test_path": "data/original/test.csv",
"processed_data_folder": "data/processed/",
"cached_model_path": "checkpoints/2019-12-15_23-25_FfnBertModelInteractor.pth",
"word2vec_path": "data/embeddings/GoogleNews-vectors-negative300.bin",
"splits": [0.85, 0.1, 0.05],
"padding": 200,
"embeddings": 1000000,
"categories": 5,
"run_model": "ffn_w2v",
"load_cached_model": false,
"models": {
"ffn_w2v": {
"data_loader_workers": 4,
"batch_size": 8192,
"learning_rate": 0.0001,
"epochs": 1,
"embedding_size": 300,
"dropout": 0.25,
"hidden": 256
},
"lstm_w2v": {
"data_loader_workers": 2,
"batch_size": 1024,
"learning_rate": 0.0001,
"epochs": 20,
"embedding_size": 300,
"dropout": 0.25,
"lstm_layers": 2,
"lstm_hidden": 128,
"lstm_dropout": 0.25,
"gradient_clip": 5
},
"ffn_bert": {
"data_loader_workers": 1,
"batch_size": 256,
"learning_rate": 0.00005,
"epochs": 2,
"embedding_size": 768,
"dropout": 0.25,
"hidden": 256,
"max_batches_per_epoch": 64
},
"lstm_bert": {
"data_loader_workers": 1,
"batch_size": 256,
"learning_rate": 0.00005,
"epochs": 1,
"embedding_size": 768,
"dropout": 0.25,
"lstm_layers": 2,
"lstm_hidden": 128,
"lstm_dropout": 0.25,
"gradient_clip": 5
}
},
"seed": 42,
"cache": true
}
See the attached license file for further notice.
I hereby explicitly prohibit using any parts of this repository with respect to the course given in 2019, tldr: don't copy my project!
If you discover this page in one of the following courses (2020+), feel free to take it as a basis in accordance to the license.