Regularization-in-Deep-Neural-Networks
The dataset has 4,844 classes, with a significant imbalance between the numbers of samples in these classes, which ranges from one sample to several hundred samples. The training split includes 341,300 segmented samples and the test split includes 1,825 samples extracted from 300 documents.
You can download the dataset and pretrained weights from the following link address. https://drive.google.com/file/d/1ccbNnXEf7NeOFWRQku4-eNaHxKA-2iBN/view?usp=sharing https://drive.google.com/file/d/1y10XYHuuAQo0HJxOe8jbVegFWiBF9s8I/view?usp=sharing
There are two folders, train and val. The subfolders in each folder represent the classes. Codes are in Tensorflow and Pytorch.