Рет қаралды 1,996
Code associated with these tutorials can be downloaded from here: github.com/bns...
The accuracy of a deep learning model depends on the amount and diversity of training data.
Typical deep learning models have millions of parameters (weights & biases) to tune.
This requires large amount of data, but often we do not have or generate more data.
Data augmentation can be used to transform (e.g., zoom, rotate, etc.) and artificially enhance & diversify existing training data.
Data augmentation does not generate new data!
Image Augmentation:
All original images are transformed every epoch and used for training.
The steps per epoch (iterations) parameter is typically set to a value = total data / batch size to ensure that the network sees entire data every epoch.
Image augmentation, despite its name is not designed to artificially enhance the amount of data but rather generate new variations of existing data.
Each epoch sees a slightly different variation of the data.
Augmentation is performed on the fly during training, but you can augment images separately and save them for future use.