Final code for Kaggle lish-moa competition.
- Using
g-
(genes) andc-
(cells) features, - PCA,
- QuantileTransform,
- VarianceThreshold,
MultilabelStratifiedKFold
instead of traditional sklearn KFold- Single NN model with only few linear layers (see /src/model.py),
SmoothBCEwLogits
loss function for training (torch.nn.BCEWithLogitsLoss
for validation)
- Prediciton clipping (
np.clip(y_hat, 0.001, 0.999)
) - Averaging results from best folds (best val_loss checkpoint from each fold)
!pip install ../input/iterativestratification/
from iterstrat.ml_stratifiers import MultilabelStratifiedKFold
(only for logging)
!pip install neptune-client