unonao / kaggle-child-mind-institute-detect-sleep-states

Kaggle Child Mind Institute - Detect Sleep States

cp kaggle_cmi/shimacos/stacking_exp028_021_from_016_192step

How to use

Docker

docker compose biuld
docker compose run --rm kaggle bash # bash に入る
docker compose up # jupyter lab 起動

# cpu with db
docker compose -f compose.cpu.yaml up -d 
docker exec -it kaggle-cpu /bin/bash

How to use

preprocess

python run/prepare_dev.py 
python run/prepare_data.py phase=train
python run/prepare_data.py phase=test

train

python run/train.py exp_name=exp007_0 split=stratify_fold_0 batch_size=32 feature_extractor=CNNSpectrogram "pos_weight=[1.0, 5.0, 5.0]"


python run/cv_train.py exp_name=exp013 "pos_weight=[1.0, 5.0, 5.0]" "features=002" # 0.7393
python run/cv_score.py exp=exp013 post_process.remove_periodicity=false  # score: 0.7401
python run/cv_score.py exp=exp013 post_process.remove_periodicity=true #  score: 0.7566
python run/cv_score.py exp=exp013 post_process.remove_periodicity=true post_process.periodicity.filter_size=10000  #  score: 0.7571
python -m run.cv_inference exp_name=exp013 model.encoder_weights=null phase=train batch_size=8 "features=002" num_tta=5  # tta=1:0.7400 → 0.7566 tta=2:0.7572→0.7706 tta=3: 0.7614→0.7725 tta=5: 0.7639→0.7741

python run/cv_train.py exp_name=exp029 "pos_weight=[1.0, 5.0, 5.0]" "features=007" # 0.7393　→　?
python run/cv_score.py exp_name=exp029 post_process.remove_periodicity=true post_process.distance=40 # 80: 0.7599
python -m run.cv_inference exp_name=exp029 model.encoder_weights=null phase=train batch_size=8 "features=007" num_tta=3 # 2:0.7615→0.7728, 5:0.7668→0.7765

python run/cv_train.py exp_name=exp044 "pos_weight=[1.0, 5.0, 5.0]" batch_size=8 "features=007" model=Spec2DCNN2DayV2 epoch=15 monitor=val_score monitor_mode=max duration=17280
python run/cv_score.py exp_name=exp044 post_process.remove_periodicity=true post_process.distance=80 #  score: 0.7744
python -m run.cv_inference exp_name=exp044 model.encoder_weights=null phase=train model=Spec2DCNN2DayV2 duration=17280  batch_size=8 "features=007" num_tta=3 # 1:0.7650→0.7744 # 2: 0.7630→0.7724 # 3: 0.7745→0.7814 # 5: 0.7749→0.7811

python run/cv_train.py exp_name=exp048 "pos_weight=[1.0, 5.0, 5.0]" batch_size=8 "features=007" model=Spec2DCNNSplit model.n_split=1 epoch=30 monitor=val_score monitor_mode=max duration=17280
python run/cv_score.py exp_name=exp048 post_process.remove_periodicity=true post_process.distance=80 #  score: 0.7804
python -m run.cv_inference exp_name=exp048 model.encoder_weights=null phase=train model=Spec2DCNNSplit model.n_split=1 duration=17280  batch_size=8 "features=007" num_tta=3 # 3:　0.7742→0.7826

python run/cv_train.py exp_name=exp054_zero_periodicity "pos_weight=[1.0, 5.0, 5.0]" batch_size=8 "features=012" model=Spec2DCNNSplit model.n_split=1 epoch=30 monitor=val_score monitor_mode=max duration=17280 datamodule.zero_periodicity=True # 0.7827
python run/cv_score.py exp_name=exp054_zero_periodicity post_process.remove_periodicity=true post_process.distance=80 #  0.7840
python -m run.cv_inference exp_name=exp054_zero_periodicity model.encoder_weights=null phase=train model=Spec2DCNNSplit model.n_split=1  datamodule.zero_periodicity=True duration=17280  batch_size=8 "features=012" num_tta=1 # tta1: 0.7841→0.7840 tta3: score: 0.7848 → 0.7847

python run/cv_train.py exp_name=exp078_lstm "pos_weight=[1.0, 5.0, 5.0]" batch_size=8 "features=012" model=Spec2DCNNSplit model.n_split=1 epoch=30 monitor=val_score monitor_mode=max duration=17280 datamodule.zero_periodicity=True decoder.dropout=0.3 feature_extractor=LSTMFeatureExtractor 
python run/cv_score.py exp_name=exp078_lstm post_process.remove_periodicity=true post_process.distance=80 #  0.7840
python -m run.cv_inference exp_name=exp078_lstm model.encoder_weights=null model=Spec2DCNNSplit model.n_split=1  datamodule.zero_periodicity=True duration=17280  batch_size=8 "features=012" feature_extractor=LSTMFeatureExtractor num_tta=3 phase=train # tta1: 0.7889   tta3: score: 0.7910 

python run/cv_train.py exp_name=exp081_mixup_short_feat14 "pos_weight=[1.0, 5.0, 5.0]" batch_size=8 "features=014"  model=Spec2DCNNSplit model.n_split=1 epoch=30 monitor=val_score monitor_mode=max duration=17280 datamodule.zero_periodicity=True decoder.dropout=0.3 downsample_rate=6 feature_extractor=LSTMFeatureExtractor augmentation.mixup_prob=0.5 "augmentation.mixup_alpha=0.5" offset=5
python -m run.cv_inference exp_name=exp081_mixup_short_feat14 model.encoder_weights=null model=Spec2DCNNSplit model.n_split=1  datamodule.zero_periodicity=True duration=17280  batch_size=8 "features=014" feature_extractor=LSTMFeatureExtractor downsample_rate=6 num_tta=3 phase=train 

python run/cv_train.py exp_name=exp085_split_drop3  "pos_weight=[1.0, 5.0, 5.0]" batch_size=8 "features=014" model=Spec2DCNNSplit model.n_split=1  epoch=30 monitor=val_score monitor_mode=max duration=17280  datamodule.zero_periodicity=True datamodule.zero_periodicity=True decoder.dropout=0.3 downsample_rate=6 augmentation.mixup_prob=0.3 "augmentation.mixup_alpha=0.5" offset=5 
python -m run.cv_inference exp_name=exp085_split_drop3 model.encoder_weights=null model=Spec2DCNNSplit model.n_split=1  datamodule.zero_periodicity=True duration=17280  batch_size=8 "features=014" downsample_rate=6 num_tta=3 phase=train # 0.8001

inference

python -m run.cv_inference exp_name=exp013 model.encoder_weights=null phase=test post_process.remove_periodicity=true batch_size=8 "features=002" num_tta=2