Data Augmentation Data Augmentation Hands-on Deep Learning v2

1. Data Augmentation

https://www.github.com/aleju/imgaug

2. Code implementation

3. Q&A

• 1. If the number and diversity of the original samples are large enough, data augmentation is not required. In reality, the data is so large that the quantity and diversity are sufficient.
• 2. In the field of financial sub-control, the data on fraud is relatively small, so data augmentation is needed.
• 3. The graph neural network is currently not easy to train, and it is still more difficult to land. – 2021-06-20
• 4. Data augmentation actually increases the bias without changing the mean of the data.
• 5. The mix-up image augmentation is effective, that is, the two images are combined together, and the labels are also combined together. Why, I don't know. mix-up diagram
     

reference

https://www.bilibili.com/video/BV17y4y1g76q?p=1