ml-template

A cookiecutter template for any ML/DL project.

Create project

If you don't have cookiecutter, install it:

pip install cookiecutter

Then, run the following command line:

cookiecutter https://github.com/nicohlr/ml-template

You will be asked for some information about your project:

• PROJECT_NAME: The name of the project folder that will be created.
• TARGET_LABEL: The label of your target column.
• METRIC: The metric that will be used to evaluate models. You must choose a metric among the list of available metrics in scikit-learn. You must choose a metric that suits your problem (don't choose a regression metric if your problem is a classification ...).

A complete project folder will then be created in current path. This project folder contains subfolders, in which you can store your data, notebooks, templates and scripts:

├── LICENSE
├── README.md              <- The top-level README
├── input                  <- You can store your data here
│   ├── train.csv
│   └── test.csv
├── models                 <- All trained models
├── notebooks              <- You can store your notebooks here
├── src                    <- Source code for use in this project.
│   ├── config.py          <- Useful variables used in scripts
│   ├── create_folds.py    <- Create a training dataset for CV
│   ├── infer.py           <- Run inference on test dataset
│   ├── models.py          <- Reference all models here
│   ├── prepare.py         <- Code for cleaning, FE, scaling ...
│   └── train.py           <- Run training
└── requirements.txt       <- All requirements of the project

Training

Start by replacing the default datasets in the input folder by your data. Then run the create_folds.py script:

python create_folds.py

This will create a train_folds.csv file in input folder. This file is the train dataset shuffled with an additionnal "FOLD" column that will be used for k-fold cross validation. Then, you can evaluate a model on the fold of your choice with the train.py script. Training is then performed on all the remaining folds.

For example, to train a model on folds 1 to 4 and evaluate its performance on the fold 0 you must run:

python train.py --fold 0 --model lr

Note: setting the --fold argument to all will evaluate sequentially on each fold.

You can use other machine learning models for training by referencing them in the models.py script. You can also add cleaning, feature engineering, categorical variables encoding and scaling steps that will be applied on the data before training by filling the functions of the prepare.py script.

Note: At each training, the resulting model is automatically saved in models folder.

Inference

When you are ready to make the last training (train on the whole training set), just run the following command:

python train.py --model YOUR_MODEL

Note: When the fold argument is not specified, the training is performed on the whole training set.

Finally, you can run inference:

python infer.py --model_path ../models/YOUR_FINAL_MODEL.bin

When inference is finished, a submission.csv file, containing predictions on test set, is dumped at the root of the projet.