/sentiment-analysis

Sentiment Analysis using Pytorch-Lightning

Primary LanguagePythonMIT LicenseMIT

Sentiment Analysis

Sentiment Analysis using Pytorch-Lightning

This is heavily inspired by Ben Trevett's work. I made this so that it becomes easier for people who are new to NLP and pytorch-lightning, to get them going with a very basic framework to get started.

With this, you can

  1. Train popular architectures to according to your hyper-parameters (on IMDb dataset for now, planning to extend)
  2. Train your own architecture. (Instructions on how this can be done will be added soon)
  3. Use one of the pre-trained model, to just get the sentiment prediction of the popular models.

List of pre-trained models available

1. Simple RNN Classifier

Test Results Simple RNN Classifier results Description

The results are not very great considering that it is a binary classification problem. It was added a baseline comparison, to see how well does the naïve architecture perform on this dataset. Although it is decently fast, it is not the fastest. Consider using other architectures for getting good predictions.

2. Bi-Directinal LSTM Classifier

BiLSTM Classifier results

Description

The best architecture in terms of test accuracy. In the further manual checking done, the results were decent. It is extremely slow, in fact it is the slowest model of all the available options in this repository. This model will be on the 17th place on IMDb Benchmark leaderboard. That is really good for a model trained for less than an hour on Google Colab GPU.

3. Fast Classifier

Fast Classifier results

Description

This model architecture strikes a perfect balance of speed and accuracy. This is the default model in all the scripts. This can be considered go to model architecture, when you want to do any sentiment-analysis task. In the further manual testing, the results were sensible "most" of the time.

4. CNN Classifier

CNN Classifier results

Description

This model architecture has a great balance of speed and accuracy. This is a super fast model. This is the fastest model on this library. If you observe closely, this model is better in accuracy than Fast Classifier and also faster than it. But this is not the default model, because this model needs a sentence at least 3 words long to work. If this is handled, this is the best possible option overall for sentiment analysis task. It is not far behind the BiLSTM Classifier, just 1.3% shy. On the IMDb Benchmark it will be on 19th place, the training took less than 10 minutes.

How to train new models with available architectures

This will give a list of all the hyper-parameters you can tweak to check and have fun.

python train.py --help

How to add your new architectures

We have added a comprehensive guide to add your own architectures here. If you liked the model you developed and want to share, the pull requests are open. :)

DEFAULTS

  1. Architecture: Average Pooling over embeddings and a linear classifier over it. This model was introduced by the paper Bag of Tricks for Efficient Text Classification. Our implementation is almost verbatim from bentrevett
  2. Optimizer: Adam with 1e-3 learning rate.
  3. Device: The code by default assumes that you are using CPU. To ue a GPU , you have to run with the flag --n_gpus=1, if you are using a GPU we suggest to use a batch size of 64.