Model/model_data/simulation_system

Simulation System

Starter Readme: Steps for pipeline:

• (WIP) Set up the training development environment

  • Change directory to this folder (simulation_system)
  • Run the following command make env PYTHON_VERSION=3.10.12
  • This will install the specified python version using pyenv and select this version as the global python version
  • It will install all training packages as specified in the training-dev.txt requirements file, including the pre-commit hooks
  • Run source .training_env/bin/activate to use this environment

• (WIP) Use Makefile to start up mock up s3 service

  • By running make init, this will run the docker-compose build and docker-compose up -d, which spins up a S3 service
  • This docker compose is running in detached mode -d, so will no output anything to the terminal

• Once the Minio service is run, you can run the training.py file to start a model build process

  • This will output a model, for a given target column, and add model name composed of some of the hyperparameters
  • An example of running this file is:
    • python3 training.py --train-filepath ./model_build_data/change_data/rdsap_full/train_validation_data.parquet --test-filepath ./model_build_data/change_data/rdsap_full/test_data.parquet
  • Outputs of the pipeline are:
    • A model directory bucket
    • A target variable prefix (i.e. RDSAP_CHANGE or HEAT_DEMAND_CHANGE)
    • A model type prefix (i.e. autogluon, tensorflow etc)
    • A model name prefix (i.e. rdsap_change_medium_quality_60_TIMESTAMP)
      • This model name is made up of target variable, quality, time spent training and timestamp
      • Within this prefix, there are three folders:
        • model
          • The model path that can be loaded in the codebase
        • deployment
          • The optimised model that can be deployed (may or maynot need this)
        • metrics
          • The metrics generatted from the model (may or may not need this as this can be contained in the registry)

• Once model build is finished, you can run the prediction.py file to generate prediction

  • By default, the prediction pipeline will select the best model based on mean absolute error from the model registry
    • This can be overwritten by specifying a model_path, which will load an alternative model
  • There are two ways of getting data into the pipeline:
    • Using the --data argument:
      • This is a JSON string which can be passed as python3 predictions.py --data '{"TOTAL_FLOOR_AREA": 1}'
        • Note the single and double quotation marks, as this affects the ingestion
    • Using the --data-path argument:
      • This can be a filepath (Can imagine that we might want to pull data from S3/ DB)
  • An example of running the file is:
    • python3 predictions.py --data-path ../simulation_system/model_build_data/change_data/rdsap_full/test_data.parquet
  • Outputs of the pipeline are:
    • prediction bucket
    • a Target variables prefix (i.e. RDSAP_CHANGE or HEAT_DEMAND_CHANGE)
    • a uprn prefix (i.e 0123456789)
    • a prediction.json
    • a metadata.json
      • This is all the metadata from the model (can change this if needed)

• NOTE: If you wish to change any settings, these are currently all in the Settings.py file

  • It will be separated out eventually but for now, it works to keep track of anything that we might want to respecify.
    • I.e. the hyperparameters for models are in here but will move into a separate configuration file

TODO:

• Structure/ MLOps:
  • Add configuration files (dev, staging, prod), including hyperparamters
  • Add precommit hooks (linters, branch names, etc)
  • Sphinx documentation
  • Sort out local mock up services
  • Sort out Model Registry
  • Sort out Data version control
  • pre-commit hooks:
    • The types of hooks that we want (safety, bandit, iso8 etc)
    • The customisations we require
  • Add sphinx documentation
• Data Science:
  • Implement a metrics class, to hold all metric
  • Rebuild metrics script (Could be a one off but good to have)
  • Determine metrics
  • Implement and test custom model (Tensorflow Decision Trees etc)
• Orchestration:
  • Lambda handler for the pipeline