mirror of
https://github.com/huggingface/lerobot.git
synced 2026-05-15 00:29:52 +00:00
CarolinePascal
82 lines
7.4 KiB
Plaintext
82 lines
7.4 KiB
Plaintext
# Video encoding parameters
|
||
|
||
When **video storage** is on, LeRobot stores each camera stream as an **MP4** file rather than saving **every timestep as its own image file**. **Video encoding compress across time**, which usually cuts **dataset size and I/O** compared to heaps of PNGs, and MP4 stays a **familiar format** for players and loaders. Incoding frames into a MP4 file is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs speed, and
|
||
optional extra encoder flags. **Many of those knobs are user-tunable** and are exposed on the dataset config as
|
||
**`dataset.camera_encoder_config`** — a nested **`VideoEncoderConfig`** (`lerobot.datasets.video_utils.
|
||
VideoEncoderConfig`) passed through **PyAV**.
|
||
|
||
You can set these parameters from the CLI with **`--dataset.camera_encoder_config.<field>`** (e.g. `lerobot-record`, `lerobot-rollout`). The same block applies to **every** camera video stream in that run. **Video storage must be on** — **`use_videos=True`** in Python APIs or **`--dataset.video=true`** (recording default); with video off, inputs stay as images and **`camera_encoder_config` is ignored.**
|
||
|
||
For **when** frames are written vs encoded (streaming vs post-episode), queues, and other top-level **`--dataset.*`** switches, see [Streaming Video Encoding](./streaming_video_encoding). For codec/size/speed experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
|
||
|
||
---
|
||
|
||
## Tuning Parameters
|
||
|
||
| Parameter | CLI flag | Type | Default | Description |
|
||
| --------------- | ----------------------------------------------- | -------------------- | ------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
||
| `vcodec` | `--dataset.camera_encoder_config.vcodec` | `str` | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, else `libsvtav1`. |
|
||
| `pix_fmt` | `--dataset.camera_encoder_config.pix_fmt` | `str` | `"yuv420p"` | Output pixel format; must be supported by the specified codec in your FFmpeg build. |
|
||
| `g` | `--dataset.camera_encoder_config.g` | `int \| None` | `2` | GOP size (keyframes every `g` frames). Emitted as FFmpeg option `g`. |
|
||
| `crf` | `--dataset.camera_encoder_config.crf` | `int \| None` | `30` | Abstract **quality**; mapped per codec in the table below (CRF, QP, `q:v`, etc.). Lower → higher quality / larger output where the mapping is monotone. |
|
||
| `preset` | `--dataset.camera_encoder_config.preset` | `int \| str \| None` | `12`\* | Video encoding speed preset; meaning depends on the specified codec. \*Unset + `libsvtav1` → LeRobot sets `12`. |
|
||
| `fast_decode` | `--dataset.camera_encoder_config.fast_decode` | `int` | `0` | `libsvtav1`: `0–2` passed in `svtav1-params`; `h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`; other codecs: often unused. |
|
||
| `video_backend` | `--dataset.camera_encoder_config.video_backend` | `str` | `"pyav"` | Only `"pyav"` is implemented for video encoding today. |
|
||
| `extra_options` | (nested config / non-scalar) | `dict` | `{}` | Extra FFmpeg options merged after the built-in mapping; **cannot** override keys already set from structured fields above. |
|
||
|
||
---
|
||
|
||
## Validation
|
||
|
||
| What | Behavior |
|
||
| -------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
||
| Video codec presence | `vcodec` must exist as a video encoder in the local FFmpeg build (after resolving `"auto"`). |
|
||
| Pixel format | `pix_fmt` is checked against the encoder’s reported pixel formats when available. |
|
||
| Options | `get_codec_options()` output (including values originating from `extra_options`) is checked against PyAV/FFmpeg option metadata (ranges, integer constraints, string choices) where applicable. |
|
||
|
||
---
|
||
|
||
## Mapping: `VideoEncoderConfig` → FFmpeg options
|
||
|
||
From **`get_codec_options()`** after `vcodec` resolution. Only fields on `camera_encoder_config` are listed here (no global thread / queue flags).
|
||
|
||
| Resolved `vcodec` | `g` | Quality from `crf` | `preset` | `fast_decode` |
|
||
| ---------------------------------------- | --- | --------------------------- | -------- | ------------------------------------------ |
|
||
| `libsvtav1` | `g` | `crf` | `preset` | `svtav1-params` includes `fast-decode=0…2` |
|
||
| `h264`, `hevc` (software) | `g` | `crf` | `preset` | `tune=fastdecode` if `fast_decode > 0` |
|
||
| `h264_videotoolbox`, `hevc_videotoolbox` | `g` | `q:v` (derived from `crf`) | — | — |
|
||
| `h264_nvenc`, `hevc_nvenc` | `g` | `rc=constqp` + `qp` ← `crf` | `preset` | — |
|
||
| `h264_vaapi` | `g` | `qp` ← `crf` | — | — |
|
||
| `h264_qsv` | `g` | `global_quality` ← `crf` | `preset` | — |
|
||
|
||
---
|
||
|
||
## `extra_options`
|
||
|
||
- Merged **after** structured options; keys **already** set by `g`, `crf`, `preset`, etc. are **not** replaced by `extra_options`.
|
||
- Values are strings or numbers as FFmpeg expects; numeric values are validated when the codec exposes option metadata.
|
||
|
||
---
|
||
|
||
## Example
|
||
|
||
```bash
|
||
lerobot-record \
|
||
--robot.type=so100_follower \
|
||
--robot.port=/dev/tty.usbmodem58760431541 \
|
||
--robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
|
||
--robot.id=black \
|
||
--teleop.type=so100_leader \
|
||
--teleop.port=/dev/tty.usbmodem58760431551 \
|
||
--teleop.id=blue \
|
||
--dataset.repo_id=<my_username>/<my_dataset_name> \
|
||
--dataset.num_episodes=2 \
|
||
--dataset.single_task="Grab the cube" \
|
||
--dataset.streaming_encoding=true \
|
||
--dataset.encoder_threads=2 \
|
||
--dataset.camera_encoder_config.vcodec=h264 \
|
||
--dataset.camera_encoder_config.preset=fast \
|
||
--dataset.camera_encoder_config.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
|
||
--display_data=true
|
||
```
|