An attempt at a Kaggle competition for seizure detection.

https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/overview

There are six patterns of interest for this competition:

• seizure (SZ),
• generalized periodic discharges (GPD),
• lateralized periodic discharges (LPD),
• lateralized rhythmic delta activity (LRDA),
• generalized rhythmic delta activity (GRDA),
• “other”. Detailed explanations of these patterns are available here.

A simple git clone will suffice

$ git clone https://github.com/BlakeJC94/kaggle-hms-brain-activity

The structure of this repo is laid out as:

.
├── artifacts    # Git-ignored dir for experiment outputs
├── data         # Links and dirs for raw/processed data
├── dataset_builders  # Scripts for processing data
├── experiments       # Directory for experiment configs/hyperparams
├── hms_brain_activity  # Core ML project code
├── scrap  # Scrap files for pre-experiment exploration and testing
├── tasks  # Scripts for ML tasks
│   ├── __init__.py
│   ├── predict.py
│   ├── stop.py
│   └── train.py
└── tests  # Unit tests

Experiments are arranged as

experiments/
├── 00_toy_classifier   # Numbered and titled experiment stem
│   ├── baseline.py     # Hyperparameters defined as a JSON-like python dictionary
│   └── __init__.py     # Configuration written as functions of hparams dictionaries
├── 01_resnet1d34
│   ├── baseline.py
│   ├── decrease_lr.py  # Hyperparmeter names appended to experiment name in clearml
│   └── __init__.py
└── 02_efficientnet_spectro
    ├── baseline.py
    └── __init__.py

First, download rye:

curl -sSf https://rye-up.com/get | bash

And once it's all configured and active, navigate to the project directory and create the environment:

$ rye sync

To launch a training job (after setting up ClearML):

$ rye run ipython -- tasks/train.py <path/to/hparams.py> [--dev-run <frac or -1>] [--offline] [--debug]

Launching a training job will create a temporary file ./.train.X.pid which contains the PID of the running training job. To terminate the job, send an INT signal to that PID (or use tasks/stop.py):

$ rye run ipython -- tasks/stop.py <.train.XX.pid>

To launch an inference job:

$ rye run ipython -- tasks/predict.py <path/to/hparams.py> <path/to/weights.ckpt>

To create a zip file to submit,

$ rye run ipython -- tasks/create_submission.py <path/to/hparams.py> <path/to/weights.ckpt>

This will create a minimal zip file with the code required to perform inference and a small run.py script. To use the output:

$ unzip submission_XXXX.zip
$ mkdir data
$ ln -s <path/to/hms/dataset> data/hms
$ rye sync
$ rye run ipython -- run.py