-
QuestionI am converting a simple pytorch model including grid_sample operator into onnx and then into TRT. I've successfully executed the conversion to both ONNX and TensorRT. However, the runtime in both ONNX and TensorRT is notably lengthy. In ONNX, when employing the CUDAExecutionProvider, I encountered warnings stating, 'Some nodes were not assigned to the preferred execution providers, which may or may not have a negative impact on performance.' And in TensorRT, I observed that the grid_sample operator is executed on the CPU. Could there be any issues with my conversion to ONNX? Alternatively, could I be using an incorrect package version?
Further informationtorch 1.12.1+cu113 |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
@ruolinsss thank you for reporting this issue. Cuda implementation of GridSample will be implemented in onnxruntime. |
Beta Was this translation helpful? Give feedback.
@ruolinsss thank you for reporting this issue. Cuda implementation of GridSample will be implemented in onnxruntime.