SwayStar123/preprocessed_commoncatalog-cc-by_DCAE
收藏Hugging Face2025-01-29 更新2025-02-15 收录
下载链接：
https://hf-mirror.com/datasets/SwayStar123/preprocessed_commoncatalog-cc-by_DCAE
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是预处理的公共目录（CC-BY）DCAE，包含了通过DC-AE f32自动编码器编码的图片和经过moondream2生成、siglip和bert编码的文本描述。图片大小经过调整，确保所有边长可被32整除以适配DCAEf32编码器。文本描述部分，文本向量填充至64个标记，并提供了未填充的长度以便在批处理中剪裁以节省计算资源。
This dataset is the preprocessed Common Catalogue (CC-BY) DCAE, which includes images encoded with the DC-AE f32 autoencoder and text descriptions generated by moondream2 and encoded with siglip and bert. The images are resized to ensure that all side lengths are divisible by 32 to fit the DCAEf32 encoder. For the text descriptions, the text embeddings are padded to 64 tokens, and the unpadded length is provided for pruning to the maximum in the batch to save computational resources.
提供机构：
SwayStar123
