Neural Translation Using Couchbase

Running Couchbase

• Couchbase can be run as a Docker container

  docker run -d -p 8091-8096:8091-8096 -p 11210:11210 --name ml_db couchbase

• Setting Up

  Access Couchbase Web Console at http://localhost:8091/

  While setting up, enable Analytics, Data, Query and Indexing Service in Couchbase.

  More information on setting up can be found here.

• Enabling Python UDFs on Couchbase

  We need to enable the developer preview mode to enable us to use the machine learning capabilities of Couchbase.

  docker exec -it ml_db bash /opt/couchbase/bin/couchbase-cli enable-developer-preview -c localhost:8091 -u <username> -p <password> --enable

Using the Pre-trained Model

• Install the requirements

  pip install -r requirements.txt

• The model used is a pre trained multi lingual model, Helsinki-NLP/opus-mt-mul-en that translates from multiple common languages into English. The model is available on hugging face.

Packaging the Model

• The machine learning model can be packaged including all the dependencies using shiv for any platform. We use Linux here for our container.

  shiv --site-packages pipeline/ -o trans-lib.pyz --platform manylinux1_x86_64 --python-version 39 --only-binary=:all: transformers torch

• Copy the package into the Couchbase Container docker cp trans-lib.pyz ml_db:/tmp/

Deploying the Model

• Log in to Docker Container running Couchbase

  docker exec -it ml_db bash

• Deploy the model into Couchbase Analytics Service

  cd /tmp

  curl -v -X POST -F "data=@./trans-lib.pyz" -F "type=python" "localhost:8095/analytics/library/Default/translib" -u <username>:<password>;

Create the User Defined Function in Couchbase Analytics

• Run the User Defined Function Definition in Couchbase Analytics Query Workbench

  USE DEFAULT;

  CREATE ANALYTICS FUNCTION translate(text) AS "trans_lib", "Translate.translate_text" AT translib;

Run Machine Learning Inference

• Call the User Defined Function in Analytics Workbench

  SELECT translate("hallo") as translation

• The output from ML model can be seen on the output window

  [
    {
        "translation": [
        {
            "translation_text": "Hello"
        }
        ]
    }
  ]
  

Running the Machine Learning Interface using Python App

There is a sample application written using Streamlit to access the inference using Python.

Running the Application

• Create a bucket like messages in Couchbase to store the insurance quotes. Also create the primary index for the bucket so that we can access the data using SQL++ queries. CREATE PRIMARY INDEX on `messages`
• The database settings are accessed from a .env file. Create a copy of .env.example & rename it to .env & add the values for the bucket, host name and credentials for your Couchbase cluster.
• Run the application streamlit run chat_app.py
• The application can be used to run inference against the model running on Couchbase.
• There is an instant translation for all the messages being sent in the system.
• Apart from this, the latest messages can be viewed in a table along with the corresponding translations.