tutorial.md

Panduan Ablation Study

Kita akan menjalankan 3 eksperimen training sekaligus menggunakan Kaggle Notebook.

Langkah Persiapan

1. Buka Template Notebook Kaggle.
2. Klik tombol titik tiga ( ⋮ ) di kanan atas, lalu pilih Copy and Edit Notebook.
3. Di dalam notebook baru Anda:
   • Buka menu Add-ons > Secrets, pastikan Wandb-Api-Key sudah terpasang dan tercentang.
   • Buka menu pengaturan session (kanan atas), pastikan Accelerator diset ke GPU T4 x2.

Eksekusi Eksperimen

Ulangi langkah "Copy and Edit" di atas agar Anda memiliki 3 tab notebook yang berjalan bersamaan.

Pada masing-masing notebook, cukup ganti kode di cell paling terakhir dengan salah satu script di bawah ini, lalu jalankan (Run All):

Notebook 1 (Baseline)

!python /kaggle/input/datasets/muhyusuf1112/template-bdc/src/advanced_train.py \
    --config /kaggle/input/datasets/muhyusuf1112/template-bdc/configs/config.yaml \
    --override \
        model.backbone='convnextv2_base.fcmae_ft_in22k_in1k' \
        data.train_dir='/kaggle/input/datasets/muhyusuf1112/fix256/train_256' \
        data.test_dir='/kaggle/input/datasets/muhyusuf1112/fix256/test_256' \
        data.img_size=256 \
        data.fold_used=1 \
        model.head_type='linear' \
        model.dropout=0.4 \
        model.freeze_backbone_epochs=0 \
        training.lr=1e-4 \
        training.layer_decay_rate=0.957 \
        training.weight_decay=0.05 \
        training.batch_size=8 \
        training.cutmix_prob=0.0 \
        training.epochs=10 \
        training.grad_accum_steps=4 \
        experiment.name='ConvNeXt_Run1_Baseline' \
        experiment.group='ConvNeXt_Ablation' \
        "experiment.tags=[ConvNeXt,Ablation,Baseline]"

Notebook 2 (LR Lebih Tinggi)

!python /kaggle/input/datasets/muhyusuf1112/template-bdc/src/advanced_train.py \
    --config /kaggle/input/datasets/muhyusuf1112/template-bdc/configs/config.yaml \
    --override \
        model.backbone='convnextv2_base.fcmae_ft_in22k_in1k' \
        data.train_dir='/kaggle/input/datasets/muhyusuf1112/fix256/train_256' \
        data.test_dir='/kaggle/input/datasets/muhyusuf1112/fix256/test_256' \
        data.img_size=256 \
        data.fold_used=1 \
        model.head_type='linear' \
        model.dropout=0.4 \
        model.freeze_backbone_epochs=0 \
        training.lr=5e-4 \
        training.layer_decay_rate=0.957 \
        training.weight_decay=0.05 \
        training.batch_size=8 \
        training.cutmix_prob=0.0 \
        training.epochs=10 \
        training.grad_accum_steps=4 \
        experiment.name='ConvNeXt_Run2_LR5e4' \
        experiment.group='ConvNeXt_Ablation' \
        "experiment.tags=[ConvNeXt,Ablation,LR_Tinggi]"

Notebook 3 (Pakai CutMix)

!python /kaggle/input/datasets/muhyusuf1112/template-bdc/src/new_train.py \
    --config /kaggle/input/datasets/muhyusuf1112/template-bdc/configs/config.yaml \
    --override \
        model.backbone='convnextv2_base.fcmae_ft_in22k_in1k' \
        data.train_dir='/kaggle/input/datasets/muhyusuf1112/fix256/train_256' \
        data.test_dir='/kaggle/input/datasets/muhyusuf1112/fix256/test_256' \
        data.img_size=256 \
        data.fold_used=1 \
        model.head_type='linear' \
        model.dropout=0.4 \
        model.freeze_backbone_epochs=0 \
        training.lr=1e-4 \
        training.layer_decay_rate=0.957 \
        training.weight_decay=0.05 \
        training.early_stopping_loss_diverge_factor=100.0 \
        training.early_stopping_min_delta=0.0 \
        training.batch_size=8 \
        training.cutmix_prob=0.5 \
        training.epochs=10 \
        training.grad_accum_steps=4 \
        experiment.name='ConvNeXt_Run3_CutMix0.5' \
        experiment.group='ConvNeXt_Ablation' \
        "experiment.tags=[ConvNeXt,Ablation,CutMix]"

jangan lupa save and run all masing2 notebook