XGBoost

The XGBoost model was trained only on 50% of data since local machine only had 32 GB of RAM and loading full dataset with all spectral indices was not possible. Scaling the training to the cloud was not performed.

Data prep

For training XGBoostClassifier model a pixel-level dataset is needed. Creation of such dataset was done by buffering the kelp segmentation masks and extracting values for those pixels. A random sample of all pixels was also added to give model a change to account for inland, corrupted, cloudy and open water pixels.

To create this dataset run:

python ./kelp/data_prep/dataset_prep.py \
    --data_dir=data/raw/train \
    --metadata_fp=data/processed/stats/dataset_stats.parquet \
    --output_dir=data/processed \
    --train_size=0.95 \
    --test_size=0.5 \
    --buffer_pixels=5 \
    --random_sample_pixel_frac=0.02

Training

python ./kelp/xgb/training/train.py
    --dataset_fp=data/processed/train_val_test_pixel_level_dataset.parquet \
    --train_data_dir=data/raw/train \
    --output_dir=mlruns \
    --spectral_indices=all \
    --sample_size=0.2 \
    --experiment=train-tree-clf-exp \
    --explain_model

Training results

The best model with all input bands and all spectral indices had public LB score of 0.5125

Sample predictions:

Inference

To run inference on test images directory using model from specific run_dir use:

python ./kelp/xgb/inference/predict.py \
    --run_dir=<run_dir> \
    --data_dir=data/raw/test/images \
    --output_dir=data/predictions

To create submission run:

python ./kelp/xgb/inference/predict_and_submit.py \
    --run_dir=<run_dir> \
    --data_dir=data/raw/test/images \
    --output_dir=data/submissions/xgb \
    --preview_submission \
    --preview_first_n=10

Feature importance

The feature importance (weight) is logged by enabling --explain_model flag in the training config.

Logged figures

• Confusion matrices

• Precision-Recall curve

• ROC curve