Use per-episode TsFile layout matching LeRobot data_path

README.md

@@ -2,7 +2,7 @@
 license: other
 language:
   - en
-pretty_name: "Vacuame/train4 (LeRobot SO-100) — TsFile"
+pretty_name: "Vacuame/train4 (LeRobot SO-100) - TsFile"
 tags:
   - time-series
   - tsfile
@@ -15,79 +15,85 @@ task_categories:
   - robotics
 ---
 
-# Vacuame/train4 (LeRobot SO-100) — TsFile
+# Vacuame/train4 (LeRobot SO-100) - TsFile
 
-This dataset converts the numeric time-series of the HuggingFace LeRobot dataset
-[`Vacuame/train4`](https://huggingface.co/datasets/Vacuame/train4)
-**losslessly to the [Apache TsFile](https://tsfile.apache.org/) format**, and keeps
-the original camera videos alongside.
+This dataset converts the numeric time-series from [`Vacuame/train4`](https://huggingface.co/datasets/Vacuame/train4) to Apache TsFile while preserving the LeRobot episode-level layout.
 
-## Original dataset
+## Repository Layout
 
-- **Source dataset**: [Vacuame/train4](https://huggingface.co/datasets/Vacuame/train4)
-- **Format**: LeRobot v2.0 (robot_type = `so100`)
-- **Content**: a small SO-100 robot-arm manipulation dataset — **119 frames /
-  2 episodes / 30 fps**, 1 task ("try"), with two RGB camera streams
-  (`laptop`, `phone`).
+```text
+README.md
+data/chunk-000/episode_000000.tsfile
+data/chunk-000/episode_000001.tsfile
+meta/info.json
+meta/tasks.jsonl
+meta/episodes.jsonl
+meta/stats.json
+videos/chunk-000/observation.images.laptop/episode_000000.mp4
+videos/chunk-000/observation.images.laptop/episode_000001.mp4
+videos/chunk-000/observation.images.phone/episode_000000.mp4
+videos/chunk-000/observation.images.phone/episode_000001.mp4
+```
+
+## Source Dataset
 
-## What is in this repository
+- Source: [`Vacuame/train4`](https://huggingface.co/datasets/Vacuame/train4)
+- LeRobot version: v2.0
+- Robot type: `so100`
+- Episodes: 2
+- Frames: 119
+- FPS: 30
+- Task: `try`
+- Video streams: `observation.images.laptop`, `observation.images.phone`
 
+The original source data is one Parquet file per episode:
+
+```text
+data/chunk-000/episode_000000.parquet
+data/chunk-000/episode_000001.parquet
 ```
-data/
-└── vacuame_train4.tsfile          # numeric time-series (converted)
-videos/
-└── chunk-000/
-    ├── observation.images.laptop/episode_0000{00,01}.mp4   # copied verbatim
-    └── observation.images.phone/episode_0000{00,01}.mp4    # copied verbatim
-meta/
-├── info.json                      # original LeRobot info + tsfile_conversion notes
-├── tasks.jsonl                    # copied verbatim
-├── episodes.jsonl                 # copied verbatim
-└── stats.json                     # copied verbatim
+
+The converted data keeps the same episode granularity:
+
+```text
+data/chunk-000/episode_000000.tsfile
+data/chunk-000/episode_000001.tsfile
 ```
 
-## TsFile storage mapping (table model)
-
-| Role | Column(s) | Type | Notes |
-|------|-----------|------|-------|
-| **TAG** | `episode_id` | STRING | `episode_{episode_index}`, 2 devices (one per episode) |
-| **Time** | `round(frame_index * 1000 / 30)` ms | INT64 (ms) | 30 fps; frame_index restarts at 0 each episode |
-| **FIELD** | `state_0` … `state_5` | FLOAT | `observation.state[6]` expanded |
-| **FIELD** | `action_0` … `action_5` | FLOAT | `action[6]` expanded |
-| **FIELD** | `episode_index`, `frame_index`, `sample_index`, `task_index` | INT64 | indices (`index` → `sample_index`) |
-| **FIELD** | `episode_timestamp_s` | FLOAT | (`timestamp`) |
-
-## Conversion notes
-
-- **Only the numeric time-series is converted**. The `observation.images.*` features
-  are videos in LeRobot — their pixels live in the MP4 files under `videos/`, which
-  are kept verbatim here.
-- **TAG = `episode_id`** (2 devices). **Time = `round(frame_index × 1000/30)` ms**.
-  Because `frame_index` restarts at 0 within each episode and is strictly increasing,
-  every device's time axis is strictly increasing — no de-duplication or offset needed.
-- **Array columns expanded**: `observation.state[6]` → `state_0..state_5`,
-  `action[6]` → `action_0..action_5` (FLOAT, matching the source float32).
-- **Column names** with dots are made TsFile-safe (`next.*` etc.).
-- **No columns dropped, no rows dropped**: all 119 frames preserved (episode 0: 87,
-  episode 1: 32). The source data is one parquet per episode; they are merged on read.
-
-## Usage
-
-```python
-from tsfile import TsFileReader
-
-reader = TsFileReader("data/vacuame_train4.tsfile")
-schemas = reader.get_all_table_schemas()
-tname = next(iter(schemas))
-
-cols = ["episode_id", "state_0", "action_0", "task_index"]
-with reader.query_table(tname, cols, batch_size=65536) as rs:
-    while (batch := rs.read_arrow_batch()) is not None:
-        df = batch.to_pandas()
-        # ... process ...
-reader.close()
+## TsFile Mapping
+
+- Table name: `vacuame_train4`
+- TAG columns: `episode_id`, `task_id`
+- Time precision: milliseconds
+- Time definition: `round(frame_index * 1000 / 30)`
+- `timestamp` is preserved as `episode_timestamp_s`
+- `index` is renamed to `sample_index`
+- `observation.state[6]` is flattened to `state_0` ... `state_5`
+- `action[6]` is flattened to `action_0` ... `action_5`
+- Row count preserved: 119 source rows -> 119 TsFile rows across 2 files
+
+## Video Alignment
+
+Videos are not stored inside TsFile. They are mirrored under `videos/` and aligned by episode and frame:
+
+```text
+data/chunk-000/episode_000000.tsfile
+videos/chunk-000/observation.images.laptop/episode_000000.mp4
+videos/chunk-000/observation.images.phone/episode_000000.mp4
 ```
 
+Within an episode, `frame_index` aligns with the video frame index, and `episode_timestamp_s` aligns with video time in seconds.
+
+## Metadata
+
+`meta/info.json` has been updated so:
+
+```json
+"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.tsfile"
+```
+
+It also includes a `tsfile_conversion` object documenting the source path, converted path, TAG columns, time mapping, row count, and vector flattening.
+
 ## Citation
 
 ```bibtex

meta/info.json

@@ -11,7 +11,7 @@
   "splits": {
     "train": "0:2"
   },
-  "data_path": "data/vacuame_train4.tsfile",
+  "data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.tsfile",
   "video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
   "features": {
     "action": {
@@ -125,37 +125,24 @@
     }
   },
   "tsfile_conversion": {
-    "data_format": "tsfile",
-    "note": "data/ converted from the original LeRobot parquet (one-per-episode, merged) to a single Apache TsFile. Videos kept verbatim under videos/.",
-    "tsfile_table_name": "vacuame_train4",
-    "time_column": "Time",
-    "time_unit": "ms",
-    "time_definition": "round(frame_index * 1000/30) ms (30 fps); frame_index restarts per episode",
+    "source_dataset": "Vacuame/train4",
+    "source_data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
+    "converted_data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.tsfile",
+    "table_name": "vacuame_train4",
+    "time_precision": "ms",
+    "time_column": "Time = round(frame_index * 1000 / 30)",
     "tag_columns": [
-      "episode_id"
+      "episode_id",
+      "task_id"
     ],
-    "tag_definition": "episode_id = \"episode_{episode_index}\", one device per episode",
-    "column_mapping": {
-      "observation.state[6]": [
-        "state_0",
-        "state_1",
-        "state_2",
-        "state_3",
-        "state_4",
-        "state_5"
-      ],
-      "action[6]": [
-        "action_0",
-        "action_1",
-        "action_2",
-        "action_3",
-        "action_4",
-        "action_5"
-      ],
-      "timestamp": "episode_timestamp_s",
-      "index": "sample_index",
-      "observation.images.laptop": "kept as video under videos/ (not in tsfile)",
-      "observation.images.phone": "kept as video under videos/ (not in tsfile)"
-    }
-  }
-}
+    "row_count": 119,
+    "episode_files": 2,
+    "vector_columns_flattened": {
+      "observation.state": "state_0..state_5",
+      "action": "action_0..action_5"
+    },
+    "frame_alignment": "episode_index selects the TsFile/video file; frame_index aligns with video frame index; episode_timestamp_s aligns with video time in seconds.",
+    "notes": "meta/ and videos/ are mirrored from the source repository; only Parquet data is converted to TsFile."
+  },
+  "source_data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+}