Technical Guide

Plan using AI Image Classifier

Metal empowers the seamless integration of Extract, Transform, Load (ETL) processes with Artificial Intelligence (AI). This synergy is achieved through predefined AI Engines, as detailed in the AI Engines documentation.

ℹ️ INFO

Presently, Metal Server exclusively supports read operations with Plan tables.

Prerequisites

Before experimenting with this use case, ensure the deployment and readiness of the sample project. Refer to the Sample project documentation for guidance.

Presentation

Imagine having a dataset of images that require classification using the TensorFlowJS engine.

The goal is to configure a plan that classifies these images and exposes the rendered data through an HTTP API.

Configuration

Begin with a minimal Metal configuration file, config.yml:

version: "0.3"
server:
  port: 3000
  authentication:
    type: local
    default-role: all
  request-limit: 100mb

In this configuration:

• Set the Metal server port to 3000/TCP with port: 3000
• Enable authentication with authentication:
• Limit the maximum response size to 100 Mbytes with request-limit: 100mb

Add a users section with the user myapiuser:

roles:
  all: crudla
users:
  myapiuser:
	  password: myStr@ngpa$$w0rd

In the ai-engines section, declare an AI engine named img-class:

ai-engines:
  img-class:
    engine: tensorflowjs
    model: image-classify

Here, we:

• Create a new instance img-class using tensorflowjs as an engine
• Use the model image-classify for image classification

Configure the plans section with the plan p-image-classify containing an entity named img2class that inserts fake data and runs the declared AI engine:

plans:
  p-image-classify:
    img2class:
      - insert:
          data:
            - img: https://thumbs.dreamstime.com/b/isolated-classical-guitar-photo-png-format-available-full-transparent-background-54363220.jpg
            - img: https://img.freepik.com/premium-photo/golden-retriever-lying-panting-isolated-white_191971-16974.jpg
            - img: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTPpDgjJjrakUj3VlDC0K6XtdCRu5TK7CPXr7QHVjiL&s
      - run:
          ai: img-class
          input: img
          output:
            class: class

Step Command	Block
- insert: data: - img: https://thumbs.dreamstime... - img: https://img.freepik... - img: https://encrypted-tbn0.gstatic...	Inserts the images in the plan data
- run: ai: img-class input: img output: class: class	Runs AI img-class on the img field and creates a new field (class) for the AI output

ℹ️ INFO

For more information about using plans, please refer to: Configuration File Reference (Section Plans).

Add the source plan-image-classify to the sources section, pointing to the plan p-image-classify:

sources:
  plan-image-classify:
    provider: plan
    database: p-image-classify

To expose the plan through an HTTP API, include a schemas section with the schema aiplan pointing to the source plan-image-classify:

schemas:
  aiplan:
    source: plan-image-classify

The final configuration will be:

version: "0.3"
server:
  port: 3000
  authentication:
    type: local
    default-role: all
  request-limit: 100mb

roles:
  all: crudla
users:
  myapiuser:
	  password: myStr@ngpa$$w0rd

sources:
  plan-image-classify:
    provider: plan
    database: p-image-classify

schemas:
  aiplan:
    source: plan-image-classify

ai-engines:
  img-class:
    engine: tensorflowjs
    model: image-classify

plans:
  p-image-classify:
    img2class:
      - insert:
          data:
            - img: https://thumbs.dreamstime.com/b/isolated-classical-guitar-photo-png-format-available-full-transparent-background-54363220.jpg
            - img: https://img.freepik.com/premium-photo/golden-retriever-lying-panting-isolated-white_191971-16974.jpg
            - img: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTPpDgjJjrakUj3VlDC0K6XtdCRu5TK7CPXr7QHVjiL&s
      - run:
          ai: img-class
          input: img
          output:
            class: class

With the configuration set, restart the Metal server:

docker compose restart metal

Playing with the HTTP API

To test the HTTP API, begin by logging in:

curl --request POST \
  --url http://localhost:3000/user/login \
  --header 'content-type: application/json' \
  --data '{"username":"myapiuser","password": "myStr@ngpa$$w0rd"}'

You should receive a response with a token:

{
  "token":"eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJ1c2VybmFtZSI6Im15YXBpdXNlciIsImlhdCI6MTcwNTA1NjMzOCwiZXhwIjoxNzA1MDU5OTM4fQ._ugUqb3BfQRWe0xYqFbvIzmlBrKMXf5EYws9aWozsaE"
}

Then, select data from the "img2class" table using the provided token after the "Bearer" prefix:

curl --request GET \
  --url http://localhost:3000/schema/aiplan/img2class \
  --header 'authorization: Bearer eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJ1c2VybmFtZSI6Im15YXBpdXNlciIsImlhdCI6MTcwNTA1NjMzOCwiZXhwIjoxNzA1MDU5OTM4fQ._ugUqb3BfQRWe0xYqFbvIzmlBrKMXf5EYws9aWozsaE' \
  --header 'content-type: application/json'

You should receive a response showing rows with the img field and a new field class containing the classification of each image:

{
  "schema": "aiplan",
  "entity": "img2class",
  "status": 200,
  "metadata": {},
  "fields": {
    "img": "string",
    "class": "object"
  },
  "rows": [
    {
      "img": "https://thumbs.dreamstime.com/b/isolated-classical-guitar-photo-png-format-available-full-transparent-background-54363220.jpg",
      "class": [
        {
          "className": "screw",
          "probability": 0.6480134725570679
        },
        {
          "className": "acoustic guitar",
          "probability": 0.31899121403694153
        },
        {
          "className": "electric guitar",
          "probability": 0.014988371171057224
        }
      ]
    },
    {
      "img": "https://img.freepik.com/premium-photo/golden-retriever-lying-panting-isolated-white_191971-16974.jpg",
      "class": [
        {
          "className": "golden retriever",
          "probability": 0.9861685633659363
        },
        {
          "className": "Labrador retriever",
          "probability": 0.009661964140832424
        },
        {
          "className": "Sussex spaniel",
          "probability": 0.0011867072898894548
        }
      ]
    },
    {
      "img": "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTPpDgjJjrakUj3VlDC0K6XtdCRu5TK7CPXr7QHVjiL&s",
      "class": [
        {
          "className": "sports car, sport car",
          "probability": 0.23274612426757812
        },
        {
          "className": "convertible",
          "probability": 0.17254509031772614
        },
        {
          "className": "pickup, pickup truck",
          "probability": 0.16245432198047638
        }
      ]
    }
  ]
}

Feel free to explore and interact with the HTTP API using the provided example.

For additional details and comprehensive information, please consult the API documentation.