nsl2014fm / imaterialist-furniture-2018

Kaggle competition

Kaggle: imaterialist-challenge-furniture-2018

This is part of code for the first place in the competition https://www.kaggle.com/c/imaterialist-challenge-furniture-2018/leaderboard

Competition: image classification with 128 classes Link: https://www.kaggle.com/c/imaterialist-challenge-furniture-2018 Result: private leaderboard score 0.12565 (to get our final score we averaged much more models)

How to run

1. Download data from kaggle to ./data/
2. Download images python downloader.py
3. Train models python cnn_runner_0.py
4. Train models python cnn_runner_1.py
5. Train models python cnn_runner_2.py
6. Train models python cnn_runner_3.py
7. Train models python cnn_runner_4.py
8. Predict python predict_all.py
9. Generate submission Submit.ipynb

Overview:

• ensemble inceptionv4, densenet161, densenet201, inceptionresnetv2, xception
• 12TTA
• probability calibration from train distribution to balanced distribution https://www.kaggle.com/dowakin/calibrate-probability-0-005-to-lb

Some results from my experiments (1070):

• resnet34 - epoch 18 val 0.62678 0.172 (15min*18epoch)
• resnet50 - epoch 4 val 0.63055 0.171 (30min*4)
• resnet101 - epoch 6 val 0.59619 0.157 (43min*6)
• inception3 - epoch 8 val 0.60509 0.154 (37min*8) = 0.16927 LB
• densenet121 - epoch 17 val 0.60620 0.167 (28min*17)
• densenet161 - epoch 7 val 0.57006 0.149 (60min*7) = 0.16406 LB
• densenet201 - epoch 11 val 0.54275 0.145 (45min*11) = 0.15755 LB
• densenet161 - epoch 13 val 0.53795 0.150 = 0.15130 LB
• inceptionv4 - epoch 15 val 0.52382 0.137 (70min*15)
• inceptionresnetv2 - epoch 11 val 0.49438 0.139 (65min*11)
• xception - epoch 14 val 0.53719 0.149 (65min*15)