ML_Workflow

Repo overview for solving Kaggle's Tweet Sentiment Extraction Kaggle's Tweet Sentiment Extraction Competition.

• Trained TensorFlow roBERTa model (from Huggingface)
• Deployed via a Flask webservice in a Docker container
• Unittests

Installation

To start the project, follow these steps:

1. Clone Repo:

   git clone https://github.com/erlbacsi/ML_Workflow.git

2. Install dependencies:

   Dependencies are all in the ‘requirements.txt’ file. This is used to build the Docker image

   docker build -t webservice_image .

   After that, the Docker container can be built to start the web service in the container. To do this, the script ‘webservice.py’ is executed

   docker run -d -p 5000:5000 --name webservice_image webservice

   After the container has been started, the web service is active and can now respond to corresponding requests.

Usage

This is how the web service can be used:

1. Prepare sentiment input

   The input is expected in JSON format. The PowerShell script ‘shell_request.ps1’ is already available for this purpose. This script contains a dictionary into which any text can be inserted after the key ‘text’. In addition, the model requires input for the key ‘sentiment’. A distinction is made between three types of text. The key terms for this are ‘positive’, ‘neutral’ and ‘negative’.

2. Send a request

   The script can then simply be started in a PowerShell.

   .\shell_request.ps1

   The script sends the text to the model and receives a corresponding response, which is then displayed in the PowerShell. In addition, the response time in seconds is displayed.

Unittests

The script ‘test_webservice.py’ is located within the Docker container and checks the web service for some test cases. To test the web service, a command line is opened on the Docker container.

 docker exec -it webservice

The script is located in the ‘usr/src/app’ folder and can simply be started via the CMD.

python test-webservice.py -v

The unit tests check the web service once for correct input and once for incorrect input. The different cases can be seen in the Python script.

Latency time improvements

• Using GPU (CUDA) for model prediction and computation
• Caching from frequent requests
• Quantisation of the ML model parameters (lower computation time and memory usage)