Compressed ImageNet
ImageNet has been one of the standard benchmark datasets that are frequently used in computer vision tasks. However, it is also notorious for its size - 10 million training samples, taking up approximately 140-160GB of storage. This makes ImageNet extremely heavy for many training pipelines.
In this article, we slim down ImageNet through sharding, making it easy to distribute and suitable for DNN training. Due to ImageNet restrictions, a copy of the processed data will not be provided in this page.