spaicom-lab/semantic-tiny-imagenet

该数据集名为Latents for tiny-imagenet (timm)，是tiny-imagenet数据集的潜在表示集合，由timm库中的多种预训练视觉模型生成。数据集包含多个配置，每个配置对应不同的模型架构（如AIMv2、BEiT、CAFormer等）和训练设置（如patch大小14、16、分辨率224、336、448等），以及不同的预训练和微调变体（如in1k、in22k表示在ImageNet-1k或ImageNet-22k上训练）。数据以parquet格式存储，分为训练集和验证集分割，适用于计算机视觉任务中的特征提取、迁移学习或潜在空间分析。

The dataset named Latents for tiny-imagenet (timm) is a collection of latent representations for the tiny-imagenet dataset, generated by various pre-trained vision models from the timm library. It includes multiple configurations, each corresponding to different model architectures (e.g., AIMv2, BEiT, CAFormer) and training settings (e.g., patch sizes 14, 16, resolutions 224, 336, 448), along with variants of pre-training and fine-tuning (e.g., in1k, in22k indicating training on ImageNet-1k or ImageNet-22k). The data is stored in parquet format, split into train and valid sets, and is suitable for feature extraction, transfer learning, or latent space analysis in computer vision tasks.