Dali: How is WarpAffine supposed to be used?
I'm trying to use the WarpAffine op. It seems like it only supports constant transformation matrix which is almost never what you want to do. How can I use WarpAffine with dynamic matrices?
I tried generating and feeding matrices through ExternalSource but am getting the following error:
"failed: Argument input may only be produced by support op."
What is a support op? Searching for "support op" on docs shows zero results. Can you provide an example usage of WarpAffine?
vahidk
All 2 comments
Hi,
Yes in the most recent release the WarpAffine operator is limited. However, it is under heavy rework and thanks to https://github.com/NVIDIA/DALI/pull/1390 or https://github.com/NVIDIA/DALI/pull/1387 it should be possible to do exactly what you are asking for.
So stay tuned and check our nightly builds.
@mzient anything to add?
JanuszL
on 19 Oct 2019
It will be released in 0.16. Now the matrices can be provided as tensor arguments (matrix = <cpu_op_output>) or even as a second regular input, In the latter case, the matrices can be passed in GPU memory. You can also specify output sizes (constant or per-sample) and use clamp-to-edge instead of constant fill color.
Refer to the example for details:
https://docs.nvidia.com/deeplearning/sdk/dali-master-branch-user-guide/docs/examples/warp.html
mzient
on 15 Nov 2019
Related issues
dhkim0225
路
4Comments
Usernamezhx
路
4Comments
simonJJJ
路
3Comments
dhkim0225
路
4Comments
ben0it8
路
3Comments
Most helpful comment
It will be released in 0.16. Now the matrices can be provided as tensor arguments (
matrix = <cpu_op_output>) or even as a second regular input, In the latter case, the matrices can be passed in GPU memory. You can also specify output sizes (constant or per-sample) and use clamp-to-edge instead of constant fill color.Refer to the example for details:
https://docs.nvidia.com/deeplearning/sdk/dali-master-branch-user-guide/docs/examples/warp.html