akrao9/JourneyDBCache

JourneyDBCache是一个预计算的WebDataset缓存，专为Boomer文本到图像训练设计。该数据集包含从JourneyDB风格的图像-字幕WebDataset分片生成的缓存DC-AE图像潜在表示和缓存Gemma文本隐藏状态，旨在消除Boomer训练中冻结编码器的计算成本。数据集包括部分缓存（由于Colab Pro+会话限制未完成完整运行），包含3,955,000个样本，分辨率512x512，图像潜在表示形状为[32, 16, 16]，文本嵌入形状为[300, 1536]，使用float16数据类型存储，格式为WebDataset .tar分片。它适用于Boomer模型的训练研究，但非完整JourneyDB训练分割。JourneyDBCache is a pre-computed WebDataset cache specifically designed for Boomer text-to-image training. This dataset contains cached DC-AE image latent representations and cached Gemma text hidden states generated from JourneyDB-style image-caption WebDataset shards, aiming to eliminate the computational cost of frozen encoders during Boomer training. The dataset includes a partial cache (the full run was not completed due to Colab Pro+ session restrictions), containing 3,955,000 samples with a resolution of 512x512. The image latent representations have a shape of [32, 16, 16], while the text embeddings have a shape of [300, 1536]. It is stored in float16 data type and formatted as WebDataset .tar shards. It is suitable for training research on Boomer models, but it is not the complete JourneyDB training split.