Speedup save_episode() by optimizing video encoding | (⚡️ Performance) |

[lokiledev]

What this does

Combine several optimizations regarding image writing and video encoding in order to reduce the time it takes to save an episode.

1. Directly call ffmpeg command instead of using pyav wrapper.
   This is the most impactful one, current implementation loads each image sequentially, which is an io bottleneck, then submits to pyav. The fix directly calls the ffmpeg command, delegating the image loading to ffmpeg which efficiently handles IO and encoding.

2. Set image compression level to 0 (instead of 1)
   This has already been improved by refactor(datasets): add compress_level parameter to write_image() and set it to 1 #2135.
   Since ssds are really fast now it's faster to dump uncompressed images on disk than to waste CPU time to compress them. I measured a smaller time to save an episode using this. The tradeoff is using more temporary storage while recording an episode.

3. Use the fastest encoder preset and enable maximum level of parallelism of av1 params

• Use preset=13 the tradeoff is
  • Use less cpu to encode
  • Decoding is faster, great for dataloader: 11.26s -> 8.42s
  • file size is a little bit bigger: 9.93 mb -> 11.30mb
  • compression quality is bit lower: PSNR 40.40 -> 39.83
• Enable maximum level of parallelism :lp=6.
  This uses potentially more ram but exploits all available cpu cores during encoding.

How it was tested

Wrote an external benchmark that saves dummy action/observation but uses real source images and measures:

• duration of dataset.save_episode()
• image compression quality (psnr & ssim)
• dataloader time (random access and sequential)
• dataset size

I didn't run the existing video benchmarks with this PR, if someone can double check that would be great.

I also wrote this to optimize offline processing, so it uses as much cpu core as possible during encoding.
It might impact real time recording.
However the current design already delays the encoding to after the end of the episode, so I consider it is acceptable to use all available resources there to speedup the encoding.

[Copilot]

Pull Request Overview

This PR refactors video encoding to use direct ffmpeg subprocess calls instead of PyAV bindings, removes an unused PIL import, updates logging statements to use %s formatting, and changes the default PNG compression level from 1 to 0 for faster image writing.

• Replaces PyAV-based video encoding with subprocess-based ffmpeg commands
• Removes unused PIL import and updates logging to use % formatting
• Changes default PNG compression level from 1 to 0

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/lerobot/datasets/video_utils.py	Refactored encode_video_frames to use subprocess ffmpeg commands instead of PyAV, removed unused PIL import, and updated logging statements to use % formatting
src/lerobot/datasets/image_writer.py	Changed default compress_level parameter from 1 to 0 in write_image function

Comments suppressed due to low confidence (1)

src/lerobot/datasets/image_writer.py:83

• Documentation is outdated. The docstring states the default value is 1, but the parameter default has been changed to 0 on line 71. Update the docstring to reflect 'Defaults to 0.'

            image, as used by PIL.Image.save(). Defaults to 1.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[lokiledev]

Note: I tested the fast-decode option but didn't find any speed-up in dataloader (both random access and sequential).
It could be removed from the api.