Model Invocation Endpoint

This module is responsible for getting model inference and details from the mlflow-server. This REST module will run inference on the trained model and will require sufficient amount of memory to load a model and run inference.

Set Up

To set up the rest invocation api, copy .env file using:

cd model-invocation-endpoint
cp env.template .env

Updated the variables in the .env file.

Run the Docker container

Use these commands to run the REST API at 127.0.0.1:80

docker build -t model-invocation-endpoint -f model-invocation-endpoint/Dockerfile
docker run -d --name model-invocation-endpoint -p 80:80 model-invocation-endpoint

The REST API can be viewed at http://127.0.0.1:80/docs#/default

Rest API docs

There are two different ways to get inference using the REST API

1. Using Raw Data

Raw data can be uploaded directly using `invocation' endpoint. You To get inference from a specific version of a specific model, we can use:

curl -X 'POST' \
  'http://0.0.0.0/model/{model_name}/{model_verson}/invocation' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "columns": [
    "string"
  ],
  "data": {raw_data},
  "format": {format of raw_data}
}'

Currently supported formats are pandas_split, pandas_record, tf-instances and tf-inputs.

To get inference from the best performing model, we can use:

curl -X 'POST' \
  'http://0.0.0.0/model/{model_name}/invocation/best' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "columns": [
    "string"
  ],
  "data": {raw_data},
  "format": {format of raw_data}
}'

To get inference from the latest trained model, we can use:

curl -X 'POST' \
  'http://0.0.0.0/model/{model_name}/invocation/latest' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "columns": [
    "string"
  ],
  "data": {raw_data},
  "format": {format of raw_data}
}'

2. Using Data Location

Another method to get inference is to give location of the data in the postgres db. The api would fetch the data from the database and return the inference results.

To get inference from a specific version of a specific model, we can use:

curl -X 'POST' \
  'http://0.0.0.0/model/{model_name}/{model_version}/metadata-invocation/' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "filename": "string",
  "classname": "string",
  "dbname": "string",
  "starttime": "2021-07-28T14:16:43.026Z",
  "endtime": "2021-07-28T14:16:43.026Z",
  "user_id": "string"
}'

Here filename and classname will be used to identify the study. The object of class classname from file filename is supposed generate queries which will be used to fetch data from the postgres db and the preprocess_data function from class classname is used to pre process the data.

In the case of lung study, the classname is LungStudy and filename is lung_study. This will fetch LungStudy class from the model_class/lung_study.py file.

dbname is the name of the database.

The desired data from user with id user_id will be fetched between time starttime and endtime. starttime and endtime are optional parameters and can be None if the desired data do not have any starttime or endtime.