Ventilator Pressure Prediction

Objective

The objective of the competition was to develop a deep learning model in order to predict the pressure of ventilators given some parameters and a control variable. The competition was organized by Princeton University and & Google Brain at the Kaggle platform.

Fig. Samples of u_in (input variable) & pressure for train and test data

Description of the solution

• Models: Bi-LSTM (4 stacked layers)
• Scheduler: cosine annealing scheduler with warm restarts

The final submission consists of an ensemble of 2 models as described aboved at different checkpoints. These checkpoints were obtained before the restart of the scheduler every 50 epochs.

Running

1.Training

python train_lstm.py --folds 0 1 2 3 4 5 6 7 8 9 10 --name model1 --dropout 0.0
python train_lstm.py --folds 0 1 2 3 4 5 6 7 8 9 10 --name model1 --dropout 0.15

2.Predict

Generate OOF predictions and test predictions for the 2 training models at different checkpoints

python predict.py --folds 0 1 2 3 4 5 6 7 8 9 10 --name model1 
                  --dropout 0.0 --checkpoints None 400 450 500 550 600
python predict.py --folds 0 1 2 3 4 5 6 7 8 9 10 --name model2 
                  --dropout 0.15 --checkpoints None 400 450 500 550 600

3. Generate Submission

For running the ensemble, change the parameter MODEL_NAMES and run ensemble.py

# Ensemble models: 
# {'model1':[None,600,550],'model2':[None,600,550,400]}
python ensemble.py

Results

Overall code structure is as follows:

Model	Public LB	Private LB
Model1 best	0.1346	0.1371
Model1 checkpoint - 600	0.1348	0.1366
Model1 checkpoint - 550	0.1355	0.1373
Model2 best	0.1352	0.1376
Model1 checkpoint - 600	0.1351	0.1375
Model1 checkpoint - 400	0.1378	0.1400
------	-----------	-----------
Median Ensemble	0.1312	0.1336