There are not many reference implementations of deploying deep learning models within the Django REST framework.

The master branch contains a single view which loads a Bidirectional LSTM + MLP model to classify single sentences as positive or negative.The dev branch contains the work-in-progress scaleable API with message queuing and separate model server.

Clone the repo:

git clone [email protected]:NickLeoMartin/sentiment_api.git
cd sentiment_api/

Head into the repo and create and activate a virtual environment:

virtualenv --no-site-packages -p python3 venv
source venv/bin/activate

Install the packages used:

pip install -r requirements.txt

To see the demo run Django lightweight development sever which is available on http://127.0.0.1:8000/ :

./manage.py runserver 

The demo was trained on single-sentence book reviews, avaliable here.

You can access the API directly, via http://127.0.0.1:8000/api/get_sentiment/:

curl --header "Content-Type: application/json" --request POST --data '{"input":"This documentation is terrible"}' http://127.0.0.1:8000/api/get_sentiment/
{"response": "Successful", "status": 200, "text": "This documentation is terrible", "sentiment_score": ["Negative"]}

Pre-trained model weights, parameters and text preprocessor are housed in trained_models/ . To load the pre-trained model:

from models.wrapper import SentimentAnalysisModel

weights_file = "trained_models/2018_06_15_10_weights.h5"
params_file = "trained_models/2018_06_15_10_params.json"
preprocessor_file = "trained_models/2018_06_15_10_preprocessor.pkl" 

sentiment_model = SentimentAnalysisModel.load(weights_file,params_file,preprocessor_file)

To predict on a new sentence:

sentiment_model.predict(["This documentation is terrible"])
['Negative']

Simply specify the path to a tab separated text file (see example here) and the training parameters, Then run:

from models.wrapper import SentimentAnalysisModel	
sentiment_model = SentimentAnalysisModel(
    dataset_path="data/training.txt", model_name="bilstm", save_model=True, word_embedding_dim=100,   word_lstm_size=100,fc_dim=100,
    fc_activation='tanh', fc_n_layers=2, dropout=0.5, embeddings=None, loss = 'binary_crossentropy', optimizer="adam",
    shuffle=True, batch_size=64, epochs=4)
sentiment_model.fit()

The model is not correctly tuned and overfits to the training data. It can't handle out-of-vocabulary words as well as negation i.e. "not good". Hyperparameter tuning, character-level embeddings and a more diverse training set are logical next steps. API will not scale i.e. add message queuing, a model server and containerize.

V1: Model pipeline + REST API

• Simple Boostrap interface
• Basic Class-based API endpoint
• Ajax call to API endpoint
• BaseModel class
• Bidirectional LSTM with word
• Vocabulary builder
• DocIdTransformer
• Model training
• Model storage
• Model wrapper
• Model prediction
• Local loading and prediction through API endpoint
• Downloading of weights from git-lfs
• Run demo from scratch
• YML config file
• Input validation, error handling, documentation etc.
• Dockerize project
• Deploy demo
• Stress test
• Jupyter Notebook to demonstrate model reasoning i.e. negation, OOV terms etc.

V2: Scaleable API

• Message Queuing: Redis
• Apache Server to load model and poll (prevent loading a new model for every view)
• Update DockerFile
• Deploy demo
• Stress test
• Testing Suite

Extensions: Model Improvements + Code Resusability

• Use Glove embeddings
• Hyperparameter tuning, training summary statistic, more comprehensive dataset
• Asynchronous prediction with Celery
• Downloading script for short-text sentiment data
• Character-level embeddings (Raw or FastText)
• Multi-task learning for entities and sentiment
• Separate models and API