Hi,
Thanks for this great repo, great help for our data generation pipeline, on our own robot!
By far, the capture process will get data one by one, each will trigger one GPU->CPU sync, it will take a long time if we want to capture multi data formats.
The nsys report:
After searching the doc (with help of gpt/gemini), it seems BasicWriter do have some nice supports: the configurable write queue and threads, the jpeg format, etc:
https://docs.isaacsim.omniverse.nvidia.com/latest/replicator_tutorials/tutorial_replicator_getting_started.html#custom-writer-and-annotators-with-multiple-cameras
And maybe some helper classes in isacclab.
But the state data dict we are using in this repo is super easy to understand and use, we can just add more key/value and write all of them to one big npy, for easier usage in training.
So, I am wondering, is there any plans to make the multimodal data generation, quicker?
Sincerely.
Hi,
Thanks for this great repo, great help for our data generation pipeline, on our own robot!
By far, the capture process will get data one by one, each will trigger one GPU->CPU sync, it will take a long time if we want to capture multi data formats.
The nsys report:
After searching the doc (with help of gpt/gemini), it seems BasicWriter do have some nice supports: the configurable write queue and threads, the jpeg format, etc:
https://docs.isaacsim.omniverse.nvidia.com/latest/replicator_tutorials/tutorial_replicator_getting_started.html#custom-writer-and-annotators-with-multiple-cameras
And maybe some helper classes in isacclab.
But the state data dict we are using in this repo is super easy to understand and use, we can just add more key/value and write all of them to one big npy, for easier usage in training.
So, I am wondering, is there any plans to make the multimodal data generation, quicker?
Sincerely.