This is a dataset of book images, but number of samples are quite small.
This is created for following blog posts.
- 日本語の記事 Qiita '小さなデータセットで良い分類機を学習させるとき'
- English blog post medium 'Train A Strong Classifier with Small Dataset, From Scratch? ImageNet Weights? Or AutoML? — Part 1'
- [Small Dataset -Train With Augmentation.ipynb](Small Dataset -Train With Augmentation.ipynb) - You can simply run this only. This has the best model.
- [Small Dataset -Ttrain Without Augmentation.ipynb](Small Dataset -Train Without Augmentation.ipynb) - What if we don't use augmentation? Check this.
- [Small Dataset -Train With fast.ai library - successful.ipynb](Small Dataset -Train With fast.ai library - successful.ipynb) - What if we use fast.ai library to train this dataset.
- [Dataset examples.ipynb](Dataset examples.ipynb) - For making dataset examples. You won't need this.
- Run
download.sh
to get externally dependent python codes. - Run Jupyter notebooks.
Result from [Small Dataset -Train With Augmentation.ipynb](Small Dataset -Train With Augmentation.ipynb):
Result from [Small Dataset -Train With fast.ai library - successful.ipynb](Small Dataset -Train With fast.ai library - successful.ipynb):
- yu4u/mixup-generator - mixup implementation for Keras.