gmongaras/CC12M_and_Imagenet21K_Recap_Highqual_512

这个数据集调整了[这个数据集](https://huggingface.co/datasets/gmongaras/CC12M_and_Imagenet21K_Recap_Highqual)中的图片大小，保持宽高比不变，所有图片的最大边不超过512像素，最大图片为512x512。此外，为了满足训练模型的需求，图片大小调整到最近的16的倍数。增加了一个存储桶列，以便在训练时高效地加载数据，避免处理填充问题，可以直接加载相同大小存储桶的数据。

This dataset resizes images from [this dataset](https://huggingface.co/datasets/gmongaras/CC12M_and_Imagenet21K_Recap_Highqual). Keeping the aspect ratio, all images are resized to be no larger than 512 on any side. So the largest image is 512x512. I also resize to the nearest factor of 16 as the model I am training requires data to be a factor of 16 pixels in the first place (8x downsample, 2x patch size). I add a bucket column to allow me to load data in the same bucket for efficient training purposes. Instead of trying to deal with padding, I can just load in data of the same size bucket.