Lerobot Visualize Dataset keeps on loading

ns69956 · June 20, 2025, 5:10pm

I am not sure why the hugging face visualize dataset keeps on loading when I try to visualize my dataset in Lerobot format.

Here is the link to the visualization: Visualize Dataset (v2.0+ latest dataset format) - a Hugging Face Space by lerobot

Have anyone faced such scenario?

John6666 · June 21, 2025, 1:18am

This might be related. It might be better to raise an issue on GitHub. It seems like the behavior of multiple libraries is involved.

github.com/huggingface/lerobot

LeRobotDataset May Eagerly Load Images, Making It Unusable for Large Datasets

opened 02:21PM - 18 Jun 25 UTC

klingmad

### System Info Hello LeRobot Team! Thank's for providing this great repositor…y and the dataset visualizer. Insofar as I understand it, LeRobotDataset always keeps the whole dataset in memory, which causes out-of-memory issues and makes it unusable with datasets that exceed the machine's memory. Are there any plans to provide tooling to "add" episodes to existing datasets, without keeping the whole dataset in memory and for lazy loading? The PyTorch DataLoader, for example, also require the LeRobotDatset to be in memory. Out-of-memory issues occur quite quickly when using datasets required to fine-tune PI0 and GR00T. For example, an image with 256x256 pixels and 3 channels, where each pixel is 1 byte, requires 256*256*3 = 196,608 bytes = 192KB. GR00T 1.5 suggests a fine-tuning configuration with a large batch size, say 128, and 20K steps. If each frame has two images, the dataset has 5.12 million images. This requires 937.5 GB which exceeds DGX H100 memory (8 * H100). The call stack is: `save_episode` saves the current episode and calls `save_episode_table` which calls both: 1. `embed_images` embeds image bytes into the table 2. ep_dataset = embed_images(ep_dataset) self.hf_dataset = concatenate_datasets([self.hf_dataset, ep_dataset]) and `hf_dataset` stays in memory. ### Information - [ ] One of the scripts in the examples/ folder of LeRobot - [x] My own task or dataset (give details below) ### Reproduction ep_dataset = embed_images(ep_dataset) self.hf_dataset = concatenate_datasets([self.hf_dataset, ep_dataset]) and `hf_dataset` stays in memory. ### Expected behavior After saving the episode to disk, the self.hf_dataset should be cleared or save only references to the parquet files, instead of accumulating the pngs.

Topic		Replies	Views
Loading a dataset doesn't actually memory map 🤗Datasets	1	927	September 4, 2023
Extremely slow data loading of imagefolder 🤗Datasets	9	2400	January 4, 2024
Loading Huge Image Dataset seems to take a lot of time 🤗Datasets	7	3729	May 16, 2022
Unable to load images 🤗Datasets	2	133	December 31, 2024
Large image dataset, feedback and advice: data viewer, task template, and more 🤗Datasets	5	915	November 22, 2022

Lerobot Visualize Dataset keeps on loading

Related topics