Small Book Image Dataset

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'

Jupyter Notebooks

• [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.

How to run the notebooks

1. Run download.sh to get externally dependent python codes.
2. Run Jupyter notebooks.

Sample visualizations

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):

---

Thanks to

• yu4u/mixup-generator - mixup implementation for Keras.