ieiriyuki / ml-engine

Automatic deployment of prediction model to Cloud Machine Learning Engine

Kaggle House Price Prediction

Predict the prices of houses in Kaggle competition

https://www.kaggle.com/c/house-prices-advanced-regression-techniques

Docker images

Container images are based on google/cloud-sdk:226.0.0.

houseprice:1.0.0

This is written in Dockerfile, which copies relevant files to /airflow/dags. To build, docker build -t houseprice:1.0.0 . This assumes everything are on a production environment. Therefore, you should do docker run --it -p 8080:8080 -p 5000:5000 houseprice:1.0.0 /bin/bash

google-cloud-sdk

Read below.

• https://github.com/GoogleCloudPlatform/cloud-sdk-docker/
• https://hub.docker.com/r/google/cloud-sdk
• http://docs.docker.jp/engine/reference/builder.html#volume

Google Cloud authentication is necessary. We will do them like below.

1. copy service-account-credential to /airflow/dags/.gcp/your-credential.json
2. export GOOGLE_APPLICATION_CREDENTIALS=/airflow/dags/.gcp/your-credential.json
3. gcloud auth activate-service-account --key-file=${GOOGLE_APPLICATION_CREDENTIALS}
4. rm -r .gcp
5. export PROJECT_ID="your-project-id" REGION="your-region" BUCKET_NAME="your-bucket-name" MODEL_NAME="your-model-name"
6. gcloud config set core/project ${PROJECT_ID}
7. gcloud config set compute/region ${REGION}

Basic configuration is done.

.py description

By default

1. src/inspect_data.py loads data/raw/new.csv and check its dtype.

2. src/alignment.py loads data/raw/old.csv, extract necessary features, alignment them, and saves data/prepared/alignmented_data.csv.

3. src/split.py loads data/prepared/alignmented_data.csv and split it into data/prepared/data_train.csv and data/prepared/data_valid.csv.

4. src/train.py loads data/prepared/data_train.csv, and train a linear regression model, which will be saved into data/pickles/model.pkl or data/pickles/${DATE}/model.pkl. This code also uses MLflow modules and track the coefficient of determinant R^2, Rooted Mean Squared Log Error, the linear model called lm in mlruns directory.

   4.1 calculate_metrics in src/evaluate.py is called by src/train.py to save metrics of performance of the model into data/profile/metrics.json.

5. src/predict.py loads data/pickles/model.pkl and data/raw/test.csv, and them make a prediction file data/prepared/my_submission.csv.

These steps are controled by Airflow jobs that is written in src/dags.py.

MLFlow and Airflow

It is supposed that applicatioins are evoked in a container.
Expose ports 5000 and 8080 for MLflow and Airflow respectively.

MLflow tutorial is https://mlflow.org/docs/latest/quickstart.html

Airflow tutorial is https://airflow.apache.org/tutorial.html

Commands below are written in misc/start_server so that you can just kick ./misc/start_server

Run below then airflow works on localhost:8080

airflow initdb
airflow webserver -p 8080 2>&1 > /airflow/logs/webserver.log &

To enable scheduled process, you have to run

airflow scheduler 2>&1 > /airflow/logs/scheduler/scheduler.log &

This is necessary for SequentialExecuter.

Run below then mlflow works on localhost:5000

mlflow ui -h 0.0.0.0 2>&1 > /airflow/logs/mlflow.log &