Sixth place solution to kaggle bengali handwritten digit recognition. Achieved 97.606%. A VGG16 architecture pre-trained on imagenet used as a baseline with hyperparameter tuning, fine tuning intermediate layers, data augmentation and test time augmentation.
The unconventional approach led to surprising results which caught Jeremy Howard's attention which he ended up tweeting about!
This architecture is implemented in Python 3.6 and Keras using Tensorflow as backend.
Tested code using:
- Ubuntu 14.04
- Python 3.6
codes: Contains codes to final submissionothers: Contains helper codes and experimental notebooks. VERY MESSY!
If you find this work useful in your research, please consider citing:
@inproceedings{zunair2018unconventional,
title={Unconventional Wisdom: A New Transfer Learning Approach Applied to Bengali Numeral Classification},
author={Zunair, Hasib and Mohammed, Nabeel and Momen, Sifat},
booktitle={2018 International Conference on Bangla Speech and Language Processing (ICBSLP)},
pages={1--6},
year={2018},
organization={IEEE}
}
You can also find it in IEEE Xplore Digital Library here