rfcx-kaggle

Kaggle Rainforest Connection Species Audio Detection

Step By Step

Step 1: Preprocessing

python preprocess.py

Step 2: Training Supervise Contrastive Learning and Fine-tune Classification

model_name = "densenet121"
version="v1"

for ((fold=0; fold<=4; fold++))
do
    echo "Start training fold ${fold}"
    python train.py train-model \
        --backbone_name ${model_name} \
        --fold_idx ${fold} \
        --saved_path "./checkpoints/${model_name}_${version}" \
        --pretrained_with_contrastive 1 \

    python train.py train-model \
        --backbone_name ${model_name} \
        --fold_idx ${fold} \
        --saved_path "./checkpoints/${model_name}_${version}" \
        --pretrained_with_contrastive 0 \
        --pretrained_path "./checkpoints/${model_name}_${version}/pretrained_best_fold${fold}.h5" \
done

Step 3: Multi-scale evaluation on each fold

model_name = "densenet121"
version="v1"

for ((fold=0; fold<=4; fold++))
do
    python evaluate.py run-multi-scale-eval \
        --backbone_name ${model_name} \
        --fold ${fold} \
        --checkpoints_path "./checkpoints/${model_name}_${version}"
done

Step 4: Multi-scale inference on test-set

model_name = "densenet121"
version="v1"

for ((fold=0; fold<=4; fold++))
do
    python prediction.py run-prediction \
        --backbone_name ${model_name} \
        --fold ${fold} \
        --checkpoints_path "./checkpoints/${model_name}_${version}"
done

Step 5: Ensemble

model_name = "densenet121"
version="v1"

python ensemble.py run_ensemble \
    --checkpoints_path "./checkpoints/${model_name}_${version}"