Child Mind Institute - Detect Sleep States

This repository is for Child Mind Institute - Detect Sleep States

Discussion url

Build Environment

1. install rye

install documentation

MacOS

curl -sSf https://rye-up.com/get | bash
echo 'source "$HOME/.rye/env"' >> ~/.zshrc
source ~/.zshrc

Linux

curl -sSf https://rye-up.com/get | bash
echo 'source "$HOME/.rye/env"' >> ~/.bashrc
source ~/.bashrc

Windows
see install documentation

2. Create virtual environment

rye sync

3. Activate virtual environment

. .venv/bin/activate

Set path

Rewrite run/conf/dir/local.yaml to match your environment

data_dir: 
processed_dir: 
output_dir: 
model_dir: 
sub_dir: ./

Prepare Data

1. Download data

cd data
kaggle competitions download -c child-mind-institute-detect-sleep-states
unzip child-mind-institute-detect-sleep-states.zip

2. Preprocess data

rye run python run/prepare_data.py -m phase=train,test

Train Model

The following commands are for training the model of LB0.714

rye run python run/train.py downsample_rate=2 duration=5760 exp_name=exp001 dataset.batch_size=32

You can easily perform experiments by changing the parameters because hydra is used. The following commands perform experiments with downsample_rate of 2, 4, 6, and 8.

rye run python run/train.py -m downsample_rate=2,4,6,8

Upload Model

rye run python tools/upload_dataset.py

Inference

The following commands are for inference of LB0.714

rye run python run/inference.py dir=kaggle exp_name=exp001 weight.run_name=single downsample_rate=2 duration=5760 model.params.encoder_weights=null pp.score_th=0.005 pp.distance=40 phase=test

Implemented models

The model is built with two components: feature_extractor and decoder.

The feature_extractor and decoder that can be specified are as follows

Feature Extractor

• CNNSpectrogram
• LSTMFeatureExtractor
• PANNsFeatureExtractor
• SpecFeatureExtractor

Decoder

• MLPDecoder
• LSTMDecoder
• TransformerDecoder
• TransformerCNNDecoder
• UNet1DDecoder
• MLPDecoder

Model

• Spec2DCNN: Segmentation through UNet.
• Spec1D: Segmentation without UNet
• DETR2DCNN: Use UNet to detect sleep as in DETR. This model is still under development.
• CenterNet: Detect onset and offset, respectively, like CenterNet using UNet
• TransformerAutoModel:
  • Segmentation using huggingface's AutoModel. don't use feature_extractor and decoder.
  • Since the Internet is not available during inference, it is necessary to create a config dataset and specify the path in the model_name.

The correspondence table between each model and dataset is as follows.

model	dataset
Spec1D	seg
Spec2DCNN	seg
DETR2DCNN	detr
CenterNet	centernet
TransformerAutoModel	seg

The command to train CenterNet with feature_extractor=CNNSpectrogram, decoder=UNet1DDecoder is as follows

rye run python run/train.py model=CenterNet dataset=centernet feature_extractor=CNNSpectrogram decoder=UNet1DDecoder