commonpool-128-dinov2-small
收藏资源简介:
CommonPool-128-DINOv2-small是一个基于WebDataset格式的衍生数据集,源自quinnlue/commonpool-256-ssl数据集。它包含经过下采样至128x128分辨率的RGB图像,并预计算了facebook/dinov2-small视觉模型的图像嵌入向量。数据集遵循非商业研究用途许可证。总规模为5,684,754个样本,其中训练集有5,664,754个样本(71个分片),验证集和测试集各10,000个样本(各1个分片)。每个样本包括三个文件:JPEG格式的128x128 RGB图像(JPEG质量95)、npy格式的DINOv2-small CLS嵌入向量(float32类型,形状为(384,))以及json格式的样本来源和哈希元数据。图像处理流程涉及将源256x256 JPEG图像解码为RGB,使用Lanczos重采样至128x128,然后以JPEG质量95重新编码。嵌入向量是基于最终存储的128x128 JPEG字节,使用标准Hugging Face图像处理器和facebook/dinov2-small模型计算得到的。该数据集适用于自监督学习、图像表示学习、特征提取等任务,可作为预训练特征或微调的基础数据。
CommonPool-128-DINOv2-small is a WebDataset-based derivative dataset derived from the quinnlue/commonpool-256-ssl dataset. It contains RGB images downsampled to a resolution of 128×128, as well as pre-computed image embedding vectors from the facebook/dinov2-small visual model. The dataset is licensed under a non-commercial research use license. The dataset comprises a total of 5,684,754 samples: the training set contains 5,664,754 samples across 71 shards, while the validation and test sets each contain 10,000 samples across 1 shard respectively. Each sample consists of three files: a 128×128 RGB image in JPEG format (with a JPEG quality of 95), a DINOv2-small CLS embedding vector in npy format (float32 type, shape (384,)), and json-format metadata containing sample source and hash information. The image processing workflow involves decoding the source 256×256 JPEG images into RGB format, resampling them to 128×128 using Lanczos resampling, and then re-encoding them at a JPEG quality of 95. The embedding vectors are computed using the standard Hugging Face image processor and the facebook/dinov2-small model, based on the finally stored 128×128 JPEG byte data. This dataset is suitable for tasks such as self-supervised learning, image representation learning, and feature extraction, and can serve as base data for pre-trained features or fine-tuning.
数据集概述
CommonPool-128-DINOv2-small 是一个派生自 quinnlue/commonpool-256-ssl 的 WebDataset 格式数据集,包含 128x128 的存储图像和预计算的 facebook/dinov2-small 图像嵌入。
来源信息
- 源数据集:
quinnlue/commonpool-256-ssl(修订版78cc60b3ba29bd3fe3bd212eb22faefc4906471d) - 许可证:非商业研究用途(继承自源数据集)
- 嵌入模型:
facebook/dinov2-small(修订版ed25f3a31f01632728cabb09d1542f84ab7b0056) - 嵌入特征:
last_hidden_state[:, 0] - 嵌入数据类型/形状:float32,
(384,)
数据规模
总行数:5,684,754
| 数据划分 | 行数 | 分片数 |
|---|---|---|
train |
5,664,754 | 71 |
val |
10,000 | 1 |
test |
10,000 | 1 |
样本内容
每个样本包含三个文件:
<key>.jpg:128x128 RGB JPEG 图像,质量设为 95<key>.npy:DINOv2-small CLS 嵌入,float32 形状(384,)<key>.json:紧凑的逐样本来源和目标哈希元数据
此外,manifest.parquet 保留了源清单元数据,并增加了目标图像字段,如 target_shard_path、target_index_in_shard、target_sha256、target_jpeg_sha256、target_width 和 target_height。嵌入仅存储在 tar 分片中,不会重复存入清单。
图像与嵌入处理流程
- 源 256x256 JPEG 图像解码为 RGB,使用 Lanczos 重采样缩放至 128x128,然后重新编码为质量 95 的 JPEG
- DINOv2 嵌入基于最终存储的 128x128 JPEG 字节计算:通过 PIL 重新打开图片,并应用
facebook/dinov2-small的标准 Hugging Face 图像处理器




