An End-to-end Deep Learning Workflow for Sentiment Analysis

Dependencies

Prefer to run this repo via docker

• Python >= 3.7
• Docker
• Node.js >= v11.14.0 (For front-end webapp UI)
• npm >= 6.9.0 or yarn >= 1.15.2 (For front-end webapp UI)

Overview

• rubixml (A pip library)

  • TextCNN (PyTorch implementation from scratch)
  • RNN (LSTM/GRU cell)
  • spaCy (Residual TextCNN)
  • BERT (Transformer)

• base-docker

  • Miniconda
  • PyTorch
  • Jupyter Notebook and Lab
  • Docker

• trainer (The main application to train models)

  • Analytics (LDA + Word Cloud)
  • Train model and predict the comments in test.txt
  • Docker

• webapi (Back-end)

  • Flask Server
  • Docker

• webapp (Front-end)

  • React-Redux

• Pipeline

  • buildkite or travis-cli

quick run

• Clone repo

  • Run git clone git@github.com:yinchuandong/sentiment-analysis.git

• Start Flask Server:

  • cd ./sentiment-analysis
  • ./start-webapi
  • (optinal) request http://127.0.0.1:5000/api/textcnn/predict, prefer postman

• Start Web UI:

  • (optional) instal yarn (preferred) or npm
  • cd ./sentiment-analysis/webapp
  • yarn install or npm install
  • yarn start or npm start
  • visit http://127.0.0.1:3000 in your browser (prefer Chrome)

• (Optional) Start Trainer

  • cd ./sentiment-analysis
  • ./start-trainer
  • visit http://127.0.0.1:8888 in your browser. You will see a Jupyter notebook page.
  • type the token abcd
  • note: if the token is wrong, please shutdown your local jupyter notebook.

rubixml

A pip python library that includes machine learning models. The reason of creating this library is to unify the training process and deployment process, which can avoid manually copy scripts. It also can easily manage version controls.

available models

• TextCNN

run unit test

cd ./sentiment-analysis/rubixml & pytest .

installation

Run cd ./sentiment-analysis/rubixml & pip install .

usage

# import model
from rubixml.torch_textcnn import TextCNNSentimentClassifier

# initialize model
model = TextCNNSentimentClassifier(embed_dim=50, lr=0.001, dropout=0.5)
model.fit('./example_data/train.txt', nepoch=1)

# predict using the last model weights
last_result = model.predict_prob(sentences)

# reset model with the best model weights
model.use_best_model()

# predict the model using the best model weights
best_result = model.predict_prob(sentences)

base docker

create a basic docker image which contains some necessary libraries, such as jupyter, pytorch, etc.

Link: https://hub.docker.com/r/yinchuandong/miniconda-torch/tags

Current Tag: v1.0.0

trainer

The main application to analyse dataset and train models

analytics

• file: ./trainer/analytics.ipynb
  • implement word cloud
  • implement LDA

model training

• file: ./trainer/train.ipynb
  • train TextCNN models using ./trainer/example_data/train.txt
  • predict scores based on the well-trained model using ./trainer/example_data/test.txt
  • save the best model to file system for deployment

webapi

A Python Flask Server (uWSGI + nginx) to host the well-trained machine learning models from trainer. It provides public access to models via http requests. (BASE_URL=http://127.0.0.1:5000/)

TextCNN/predict

• Endpoint

  POST /api/textcnn/predict

• Request

  name	type	description
  text	list of strings	the text to be predicted

• Response

  name	type	description
  score	list of floats	the predicted scores between 0 to 1
  status	string	indicate whether the request is successful

• Example:

webapp

A simple web page that allows user to type text on browser and show the prediction results. (BASE_URL=http://127.0.0.1:3000/)