affine_grid and grid_sample operators merge/accelleration #104296
Description
🚀 The feature, motivation and pitch
Hi,
To warp some data according to a (batch) of affine transformations, two functions called sequentially need to be used:
- affine_grid to calculate the transformed coordinates followed by
- warp_sample to do the actual warping.
Would it make sense to have also the option of a function that does the two operation in a single pass rather than having to call the two sequentially?
The advantage would be only for speed: warping according to an affinity transformation would avoid to store the result of affine_grid (shape (N,H,W,2) for a spatial warp), but the transformed coordinates could be calculated locally and then used immediately to warp the input data.
I am exploring the possibility of working on the topic and wanted to know if a contribution in this direction could be useful.
Thanks in advance for your feedback
Alternatives
No response
Additional context
No response