jzpeng's picture
Add TsFile (converted from devikaskumar/test_v4l2_fix)
64b9c43 verified
|
Raw
History Blame Contribute Delete
3.89 kB
---
license: apache-2.0
task_categories:
- robotics
tags:
- tsfile
- timeseries
- tabular
- lerobot
modality:
- timeseries
- tabular
pretty_name: devikaskumar test_v4l2_fix TsFile
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data/devikaskumar_test_v4l2_fix.tsfile
---
# devikaskumar/test_v4l2_fix (TsFile)
This repository is an Apache TsFile conversion of the LeRobot dataset
[`devikaskumar/test_v4l2_fix`](https://huggingface.co/datasets/devikaskumar/test_v4l2_fix).
Modalities: Time-series, tabular.
## Source dataset
- **Author:** `devikaskumar`
- **License:** Apache-2.0
- **Robot type:** `so101_follower`
- **LeRobot codebase:** `v2.1`
- **Task:** `test` (one task, `task_index=0`)
- **Split:** `train` (`0:52` in the source metadata)
- **Scale:** 52 episodes, 15,702 frames, one chunk, 30 fps
- **Source frame files:** `data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`
The original dataset contains 104 camera videos (two streams per episode). The
videos are not copied into this TsFile dataset. They remain available in the
source repository under
[`videos/`](https://huggingface.co/datasets/devikaskumar/test_v4l2_fix/tree/main/videos),
with files at
`videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`.
The two source video features are `observation.images.webcam` and
`observation.images.realsense`; both are 480×640 RGB, AV1, 30 fps, with no audio.
## Converted files
- **TsFile:** `data/devikaskumar_test_v4l2_fix.tsfile`
- **Table:** `devikaskumar_test_v4l2_fix`
- **Rows:** 15,702
- **Devices:** 52 episode/task tag combinations
- **TsFile size:** approximately 386 KB (the source Parquet shards total approximately 913 KB)
- **Metadata:** `meta/` is retained from the source; `meta/info.json` describes
the converted TsFile and records the source/video mapping in
`tsfile_conversion`.
## Schema
`Time` is `round(timestamp * 1000)` in milliseconds and restarts at zero for
each episode. The redundant source `timestamp` field is dropped.
TAG columns (the TsFile table/device dimensions):
- `episode_index` (`INT64` source value, stored as a TAG)
- `task_index` (`INT64` source value, stored as a TAG)
FIELD columns:
- `frame_index` (`INT64`)
- `sample_index` (`INT64`), renamed from source `index`
- `action_0``action_5` (`FLOAT`), flattened from `action[6]`
- `observation_state_0``observation_state_5` (`FLOAT`), flattened from
`observation.state[6]` (`.` becomes `_`)
## Conversion and compression
The generic LeRobot converter was run in merged mode, so all source episode
shards are stored in one table-model TsFile. Query an episode by filtering the
`episode_index` and `task_index` TAG columns. Numeric encoding follows the
compact profile requested for this conversion: FLOAT/DOUBLE use `GORILLA`,
INT32/INT64 and `Time` use `TS_2DIFF`, and physical columns use `LZ4`.
No source rows or numeric fields are intentionally removed except the redundant
`timestamp` field. The video fields are omitted from TsFile because videos are
external binary streams; their source paths and frame alignment metadata are
recorded in `meta/info.json`.
## Validation
The generated TsFile was opened and queried with the Apache TsFile Python SDK.
The authoritative TsFile metadata row count and staged Parquet row count both
equal **15,702**; query readback also returned 15,702 rows across 52 devices.
## Usage
```python
from tsfile import TsFileReader
path = "data/devikaskumar_test_v4l2_fix.tsfile"
reader = TsFileReader(path)
table = reader.get_all_table_schemas()["devikaskumar_test_v4l2_fix"]
columns = [
c.get_column_name()
for c in table.get_columns()
if c.get_column_name() != "Time"
]
with reader.query_table(
"devikaskumar_test_v4l2_fix", columns, batch_size=65536
) as result:
batch = result.read_arrow_batch()
```