This code was used to train Xception model for Kaggle competition Cdiscount’s Image Classification Challenge. In principle, the competition was about standard image classification, however following points made it difficult:
- Large number of categories (5720)
- Big training data (58.2 GB, 7 million products each having up to 4 images)
- A lot of difficult categories (different styles of books and CDs)
- Variable number of images (1-4) for each product
Description of methods
The report descibing the methods is available here.
The final results of the competition are available here. The Xception model (team name "Ardi Loot") trained with this code got 64th position (bronze) out of 627 (accuracy 0.72582).