aiforsea-telematics-safety's Introduction

AI For SEA - Safety

This project is for "AI For SEA" Challenge.

Setup Data

The data files should be structured in this format.

./data
    ./features
        [filename1].csv
        [filename2].csv
    ./labels
        [filename2].csv

To setup the data provided by Grab, run ./setup-1-dataset. You could cancel and re-run if it is stuck. The download will continue.

Setup Environment

TLDR, you could use ./setup-2-environment to setup, but you still need to activate the environment yourself. We need to make an isolated environment and install the requirements. For manual setup:

Create virtual environment with python 3 either with virtualenv or conda
- virtualenv -p python3 venv, then source venv/bin/activate
- or, conda create -n venv python=3.6, then source activate venv
Run pip install -r requirements.txt

Notebook

To view the notebook without re-running the script, open 2019-06-09+Telematics+Safety+Analysis.html file in project root.

To re-run the notebook:

Setup data in ./data directory.
Setup environment.
Open notebook with jupyter notebook, or other tools that support .ipynb
Re-run all the cells.

Train, Predict and Evaluate Model

Setup data in ./data directory.
Setup environment.
Make sure to put your test data in ./data-test directory with structure mentioned above.
Run ./model-predict -d "./data-test" -m "./model/safety_model_cnn_rf_stack.mdl" -o "./output/test_prediction.csv" to predict the test data. Run ./model-predict -h for more information.
To evaluate, use ./model-evaluate [prediction_csv_file_path] [test_label_csv_file_path]. Example: ./model-evaluate output/test_prediction.csv data-test-label/sample-labels.csv
(Optional) If you want to re-train the model, run ./model-train cnn-rf-stack -v 0.3 -s 6000 -d "./data" -m "./model/safety_model_cnn_rf_stack.mdl" to train model with 70% data and validate with 30% data. Add -s 6000 parameter to only use 6000 bookings for training and validation.

Recommend Projects