Python GPU Retrieval Examples Request

Software Python
1

I’m excited to leverage the ability to keep camera data on the GPU when using the python API.

Added CUDA stream input to Camera::retrieveImage and Camera::retrieveMeasure for improved performance in GPU asynchronous workflows.
Added GPU-optimized functions, blobFromImage, and blobFromImages, for converting images to Deep Learning model tensor inputs.
Added utility functions, Mat::convertColor, for common color conversions, such as swapping red and blue channels and removing the alpha channel.

Can you please provide a python example of getting the data into a CuPy array and ideally a pytorch tensor while keeping the data on GPU?

Related: Will the api reference docs be updated with 5.0 information soon?

2

This is still an experimental feature. Examples and full API documentation will be available soon.

3

Sounds good. For now I’ll proceed with using

bgr_mat = sl.Mat(res.width, res.height, sl.MAT_TYPE.U8_C4)
zed.retrieve_image(
                bgr_mat,
                sl.VIEW.LEFT,
                sl.MEM.GPU,
                sl.Resolution(res.width, res.height),
 )
torch.as_tensor(bgr_mat.get_data(sl.MEM.GPU), device='cuda'))
4

Hi @robots,

Thanks for the interest with the GPU retrieval!

We are currently updating our sample to showcase CuPy usage (and the performance gains it brings). It should be available in a following release.

In the meantime, your plan sounds good !

If working with a runtime critical environment, an alternative would be to use torch.from_dlpack(cupy_arr.toDlpack()) instead of torch.as_tensor(cupy_arr, device='cuda')) to profit from zero-copy data exchange (see Interoperability — CuPy 13.5.1 documentation for more info)

If while using CuPy, you faced any difficulties or anything you’d like to share, we’d be happy to adapt our docs in consequence.

1 Like