readme_backend.rst

Nirdizati Research backend

Master

Development

Django backend server for machine learning on event logs.

Running in a new environment

The docker build is available @ https://hub.docker.com/r/nirdizatiresearch/predict-python/ in any case if you prefer to setup your environment on your own you can refer the Dockerfile.

Docker Compose

On first run to setup the database, you can run:

docker-compose run server python manage.py migrate

To run the project:

docker-compose up redis server scheduler worker

To access a generic remote Django server you can use the ssh tunneling functionality as shown in the following sample:

ssh -L 8000:127.0.0.1:8000 <user>@<host>

Run an instance of the project

If you are familiar with docker-compose the docker-compose file is available, otherwise if you use PyCharm as IDE run the provided configurations. Finally, from the command line you can use the following sample commands to interact with our software.

Start server with

python manage.py runserver

Run tests with one of the following

python manage.py test
./manage.py test

NB: always run a redis-server in background if you want your server to accept any incoming post requests!

Start by running migrations and adding sample data

python manage.py migrate
python manage.py loaddata <your_file.json>

Start jobs from command line

curl --request POST \
  --header 'Content-Type: application/json' \
  --data-binary '{
    "type": "classification",
    "split_id": 1,
    "config": {
      "encodings": ["simpleIndex"],
      "clusterings": ["noCluster"],
      "methods": ["randomForest"],
      "label": {"type": "remaining_time"},
      "encoding": {"prefix_length": 3, "generation_type": "only", "padding": "zero_padding"}
    }
  }' \
http://localhost:8000/jobs/multiple

Creating a single split options.

• $SPLIT_TYPE has to be one of split_sequential, split_random, split_temporal, split_strict_temporal. By default split_sequential.
• test_size has to be from 0 to 1. By default 0.2

curl --request POST \
 --header 'Content-Type: application/json' \
 --data-binary '{
   "type": "single",
   "original_log": 1,
   "config": {
     "test_size": 0.2,
     "split_type": $SPLIT_TYPE
   }
 }' \
http://localhost:8000/splits/

Advanced configuration

Prediction methods accept configuration for sklearn classification/regression methods. The Job config must contain a dict with only the supported options for that method. The dict name must take the format "type.method". For classification randomForest this would be classification.randomForest. Advanced configuration is optional. Look at jobs/job_creator.py for default values. For example, the configuration for classification KNN would have to be like:

curl --request POST \
  --header 'Content-Type: application/json' \
  --data-binary '{
    "type": "classification",
    "split_id": 1,
    "config": {
      "encodings": ["simpleIndex"],
      "clusterings": ["noCluster"],
      "methods": ["knn"],
      "classification.knn": {
        "n_neighbors": 5,
        "weights": "uniform"
      },
      "label": {"type": "remaining_time"},
      "encoding": {"prefix_length": 3, "generation_type": "up_to", "padding": "no_padding"}
    }
  }' \
http://localhost:8000/jobs/multiple

Labelling job

Log encoding and labelling can be tested before prediction. It supports all the same values as classification and regression jobs but the method and clustering.

curl --request POST \
  --header 'Content-Type: application/json' \
  --data-binary '{
    "type": "labelling",
    "split_id": 5,
    "config": {
      "label": {"type": "remaining_time"},
      "encoding": {"prefix_length": 3, "generation_type": "up_to", "padding": "no_padding"}
    }
  }' \
http://localhost:8000/jobs/multiple

Contributors

• @stebranchi Stefano Branchi
• @dfmchiara Chiara Di Francescomarino
• @TKasekamp Tõnis Kasekamp
• @mrsonuk Santosh Kumar
• @fmmaggi Fabrizio Maggi
• @WilliamsRizzi Williams Rizzi
• @HitLuca Luca Simonetto