Kaggle Competition

Statoil/C-CORE Iceberg Classifier Challenge

Ship or iceberg, can you decide from space?

https://www.kaggle.com/c/statoil-iceberg-classifier-challenge

References

• Grover's notebook: https://github.com/groverpr/deep-learning/blob/master/image_comp_1_2.ipynb
• https://www.kaggle.com/devm2024/keras-model-for-beginners-0-210-on-lb-eda-r-d

---

To Do

• add incidence angle
• cross validation
• stacking of models (ensembling)
• look at, within various models, which images are being misclassified
• RS: notebook had CV https://github.com/irshadqemu/Kaggle-Competitions/blob/master/Planet_amazon_resnet34.ipynb

Done

• RS: running Grover's NB, with resnext50, increase image size sequentially (my notebook 7_5_xxx)
• RS: submit that notebook where I had 0.2258

---

Models

1) Grover's Baseline: resnet18

• valid log_loss = 0.2918
• test log_loss = 0.2932
• bs=32, image sz=75, resize 1.3x

---

2) resnext50

• valid log_loss = 0.269
• test log_loss = 0.2566
• notebook: http://localhost:8888/notebooks/kaggle_iceberg/7_1_resnext50_more_epoch_submitted.ipynb
• submitted Jan 14
• bs=32, sz=150, resize 1.3, sgdr 5 epochs
• Leaderboard: 2452 of 3380 --> 78 percentile

---

3) resnext50

• valid log_loss = 0.2508
• test log_loss = 0.2394
• notebook: 7_6_x
• submitted Jan 20
• bs=32, sz=75, resize 1.3, sgdr 5 epochs
• Leaderboard: 2333 of 3309 --> 70.5 percentile

---

4) resnext50

• valid log_loss = 0.2272
• test log_loss = 0.2073
• notebook: http://localhost:8888/notebooks/my_repos/kaggle_iceberg/7_7_resnext50_more_epoch_update_sz_zm15.ipynb
• submitted Jan 20
• bs=32, sz=75, resize 1.5, sgdr 5 epochs, lr=0.0075, dropout=0.55
• Leaderboard: 1997 of 3310 --> 60th percentile

---

5)

• valid log_loss = 0.xxxx
• test log_loss = 0.xxx
• notebook: 7_x
• submitted
• bs=32, sz=150, resize 1.3, sgdr 5 epochs
• Leaderboard: 2xxx of xxxx

---

test ??

• got 0.2258 ? resnext50, more epochs?

---

Kaggle Rankings

• 1997 of 3310 --> 60th percentile 0.2073
• 1655 of 3310 --> 50th percentile 0.1865
• 984 of 3310 --> 29.7 percentile 0.1498
• 827 of 3310 --> 25th percentile 0.1463
• 19 of 3310 --> 0.6 percentile 0.1002