Cited 0 time in webofscience Cited 0 time in scopus

최적의 data augmentation을 조합하여 강력한 classifier 만들기: A Kaggle competition

Title
최적의 data augmentation을 조합하여 강력한 classifier 만들기: A Kaggle competition
Translated Title
Mixing optimal data augmentations, the way to make powerful classifier: A Kaggle competition
Authors
이해찬조국래손종욱
DGIST Authors
조국래손종욱
Issue Date
2019-11-15
Citation
2019 대한임베디드공학회 추계학술대회, 167-170
Type
Conference
ISBN
9788996655312
Abstract
Car is a common transportation around us. If the car can be classified correctly, it can be useful for criminal investigation and commercial using. In this context, the Kaggle Korea community inclass competition was held. Dataset have a total of 196 classes for car, 9990 train images and 6150 test images. Our solution used the methods introduced recently in the augmentation, scheduler and optimizer fields, which allowed us to achieve highest F-1 score in competition. We used Mixup, Cutout and augmentation policy found by Auto-Augment in CIFAR-10. As a main scheduler, SuperConvergence, a kind of cyclical learning rate scheduler, was used. We chose AdamW and SGD with momentum as the optimizer and trained our models. Models were EfficientNet-B5, B6, B7, PNASNet-5-large and NASNet-A-Large. The ensemble methods we used were stratified K-Fold Cross-Validation and Soft-Voting ensemble. In the final ensemble prediction, The Test-time augmentation method was used to get better generalized predictions of classification result. Using the above methods and steps, the car can be accurately classed and we obtained the model which could be applied directly in the field. This solution proposed in paper recorded the 1st score of the Kaggle Korea community inclass competition
URI
http://hdl.handle.net/20.500.11750/14156
Publisher
대한임베디드공학회
Related Researcher
Files:
There are no files associated with this item.
Collection:
Division of Electronics & Information System2. Conference Papers


qrcode mendeley

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE