A web-app that helps to visualize a word-by-word breakdown of how sentiment analysis classifies text

• Research and decide on a machine learning model/architecture
• Pick out 2-3 datasets we can use to train
• Build a training pipeline
• Train and implement the model
• Serve the model using BentoML as an API
• Create a web app to take in input and visualize the output

Our endpoint is at https://sentiment-classifier-gy7t3p45oq-uc.a.run.app/ Our prediction endpoint can be accessed through making a POST request to https://sentiment-classifier-gy7t3p45oq-uc.a.run.app/predict.

# e.g. 
curl -X POST "https://sentiment-classifier-gy7t3p45oq-uc.a.run.app/predict" \
     -H "accept: */*" -H "Content-Type: application/json" \
     -d "{\"text\":\"Some example text.\"}"

Basically, make sure to set the content type to JSON and send a JSON in the format

{
  "text": "content"
}

If successful, you should get a 200 OK status and a body with something along the lines of [[0.8614905476570129], [0.7018478512763977], [0.617088258266449]] where each entry represents the sentiment from 0 (negative) to 1 (positive) of each word.

Currently, we have only implemented a training pipeline for the IMDB dataset but this is subject to change in the future. You can train a new classifier on the dataset by doing

python train.py

This will replace the current model in /model. model.json stores the model architecture, weights.h5 stores trained weights, and tokenizer.json stores word indices.

BentoML helps us to easily serve our Keras model through an API. You can package a new API by running

python bento_service_packager.py
> ...
> [0.07744759]
> [0.1166597 ]
> [0.18447165]
> [0.20329727]
> [0.24308157]
> [0.25030023]]
> _____
> saved model path: /Users/jzhao/bentoml/repository/SentimentClassifierService/20200604214004_F641D2

If you'd like to save the packaged API, just copy the contents into /bento_deploy

cp -r /Users/jzhao/bentoml/repository/SentimentClassifierService/20200604214004_F641D2/* bento_deploy
# or whatever the autogenerated URI is

There are a few dependency nuances to be aware of before building the actual Docker image. To make sure the build doesn't error out, edit bento_deploy/requirements.txt is

tensorflow==2.1.0
sklearn
bentoml==0.7.8

Then, we can build and push and run the image as follows

docker build -t bento-classifier:latest .
docker run -p 5000:5000 bento-classifier:latest

Then, visit localhost:5000 to see the BentoML server!

> model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
embedding (Embedding)        (None, 100, 64)           320000
_________________________________________________________________
lstm (LSTM)                  (None, 100, 64)           33024
_________________________________________________________________
dropout (Dropout)            (None, 100, 64)           0
_________________________________________________________________
lstm_1 (LSTM)                (None, 64)                33024
_________________________________________________________________
FC1 (Dense)                  (None, 256)               16640
_________________________________________________________________
dropout_1 (Dropout)          (None, 256)               0
_________________________________________________________________
out_layer (Dense)            (None, 1)                 257
_________________________________________________________________
activation (Activation)      (None, 1)                 0
=================================================================
Total params: 402,945
Trainable params: 402,945
Non-trainable params: 0
_________________________________________________________________

• 85% / 15% train-test split
• dataset is balanced (25k positive, 25k negative)
• RMSProp with 1e-3 Learning Rate and early stopping with patience of 2 epochs
• preprocessing
  • to lowercase
  • removed punctuation
  • removed <br /> tags
  • tokenized with vocab size of 5k
  • max sequence length of 100
• achieved 82.2% accuracy