# coaf_dataset_24_25 与 `coaf_dataset` 同结构、同 5000 条 episode 划分(复用 `coaf_dataset/splits/train_5k.json`),但时间采样不同: | 数据 | 帧数 | 采样索引 | |------|------|----------| | `raw/.../rgb/` | **24** | `linspace(0, T-1, 24)` | | `raw/.../rgb_align/` | **25** | `linspace(0, T-1, 25)`,与 reason 对齐 | | `state` / `action` | **(25, 7)** | 与 `rgb_align`、depth、pose、flow、follow **相同** `reason_indices` | | depth / pose / flow / follow | **25** | 由 `rgb_align` 或 25 帧 state 生成 | 每条 episode 的 `manifest.json` 记录 `rgb_indices` 与 `reason_indices`,便于核对对齐。 ## 目录结构(与 coaf_dataset 对齐) ``` coaf_dataset_24_25/ raw/episode_XXXXXX/ rgb/ # 24 帧,I2V 目标视频 rgb_align/ # 25 帧,供 depth/flow 等 reason 输入 state/ action/ instruction/ manifest.json action/action_25_for_{pose,follow,flow,depth}/ state/state_25_for_{pose,follow,flow,depth}/ modalities/{depth,flow,pose,follow}/ composed/ scripts/ ``` ## 流水线 ```bash # 1. 从 Bridge V2 TFDS 提取 raw(CPU,约数小时) python scripts/preprocess_raw.py --skip-existing # 2. 复制 state/action 到各 reason 目录 python scripts/generate_action_state_25.py # 3. 各 reason 模态(GPU,顺序或并行提交) python scripts/batch_depth_5k.py python scripts/batch_flow_5k.py python scripts/batch_pose_5k.py python scripts/batch_follow_5k.py # 依赖 coaf_dataset/modalities/masks # 4. 拼接训练视频(例:depth + rgb) python scripts/compose_all.py --version v4_depth_rgb ``` `follow` 仍使用父数据集 `coaf_dataset/modalities/masks` 中的 object mask。 ## 与 coaf_dataset 的差异 | | coaf_dataset | coaf_dataset_24_25 | |--|--------------|-------------------| | RGB | 25 | **24** | | reason 视频 | 多为 24(compose 时再采样) | **原生 25** | | state/action(raw) | 24;另用 TFDS 重采样 25 给 depth | raw 直接 **25**,与 reason 同索引 | | depth 输入 | `rgb/` 25 帧 | `rgb_align/` 25 帧 | Compose 默认:`--cond-frames 25 --rgb-frames 24`(例如 v4:25 depth + 24 rgb = 49 帧)。