Quick Check Stress Detection

This is an API that takes training data from Dreaddit database for mental health detection in texts. The idea is to train a NLP model to help supplement mental health apps with stress detection through user's texts inputs. The API is meant to be simple as it only have one function, to give prediction on whether the user is stressed or not based on the text input.

Prerequisites

MacOS or Linux
Python 3.6.13
conda
pip
Docker (For hosting)

Setting Up

This project uses Python 3.6.13 but it should work with Python 3.7 as well. The project uses Tensorflow 1.15 therefore it wouldn't work with Python 3.8 and above since the only tensorflow packages available are v2 and above.

To setup the environment, please follow these steps:

Create a new conda virtual environment in local or cloud services

conda create -n quick_check python=3.6 anaconda
conda activate quick_check

Clone the github repository

git clone https://github.com/silverkip/quick-check.git

Install the required packages in the conda environment

cd quick-check/build
conda install pip
pip install -r requirements.txt

Install TensorFlow Model Server (Optional)

Install TensorFlow Model Server (For Debian based distro). Since this project uses Tensorflow 1.15, it's better to use same version for serving.

wget 'http://storage.googleapis.com/tensorflow-serving-apt/pool/tensorflow-model-server-1.15.0/t/tensorflow-model-server/tensorflow-model-server_1.15.0_all.deb'
dpkg -i tensorflow-model-server_1.15.0_all.deb

Hosting the REST API Server

The model file can be downloaded from this link: https://quick-check-model.s3.ap-northeast-1.amazonaws.com/model.zip

Hosting with Docker (Recommended)

First of all, pull the latest TensorFlow serving Docker image with this command

docker pull tensorflow/serving

The model file is hosted and is publicly available in this S3 bucket (s3://quick-check-model) and can be easily deployed through docker with these commands

docker run \
    -p 8500:8500 \
    -p 8501:8501 \
    -e AWS_REGION=ap-northeast-1 \
    -e S3_ENDPOINT=s3.ap-northeast-1.amazonaws.com \
    -e TF_CPP_MIN_LOG_LEVEL=3 \
    tensorflow/serving
    --model_config_file=s3://quick-check-model/config/models.config

Hosting local model with Docker

To host the model locally with Docker, run these commands. Replace [Path to Model Directory] with the absolute path of the model directory.

docker run -t --rm -p 8501:8501 \
    -v "[Path to Model Directory]:/models/quick_check" \
    -e MODEL_NAME=quick_check \
    tensorflow/serving

Hosting it locally with Tensorflow Model Server

tensorflow_model_server --port=8500 --rest_api_port=8501 --model_name=quick_check --model_base_path=/model

Using the API

API Link (Hosted on Heroku) : https://quick-check-app.herokuapp.com/v1/models/quick-check:predict Since the SavedModel doesn't include preprocessing, we have to do it before sending the POST request to the server. The request_sample files provided shows two samples on how to make a request to the API server in Python and Node.js. These files also include a function that does the preprocessing.

API's Input Structure

The API takes JSON input in this specific format.

{
    "signature_name":"serving_default",
    "instances":[ {"input_ids": [...],
                   "input_mask": [...],
                   "segment_ids": [...],
                   "label_ids": ...} ]
}

Do note that the key "instances" contain a single item list containing the dictionary.

API's Output Structure

The API returns JSON in this specific format. "probabilities" contain an array with the probability vectors used to predict the label using argmax. "labels" is the prediction represented by either 0 or 1. A value of 0 means it's "Non-stress" and a value of 1 is "Stress".

{
    "predictions": [ {"probabilities": [Array],
                      "labels": ... } ]
}

Inside the "predictions" key is also a single item list containing the dictionary.

Future Improvements

A very important improvement that can be made is the addition of built-in Tokenizer into the API so that raw string can be posted to the API without having to go through pre-processing.

silverkip/quick-check