When I create a TensoRT engine from a UFF, ONNX, or some other serialized graph, plugins are automatically created for layers such as Non Max Suppression. This plugin’s parameters (such as DetectionOutputParameters) are fetched from the input UFF and the plugin object is created. However, in my code I have no way to retrieve these parameters, therefore I have to hard-code the values such as keepTopK in my code. Any changes to these in the model will require code changes. In order to improve portability, it would be really nice if we were able to find the plugin in the graph, get its parameters, and then be able to use these parameters in our program. That way, we could change our model without having to change our code. Hope NVIDIA considers this for upcoming releases and please comment if you’d find this feature useful.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| TensorRT Plugin data access | 1 | 332 | April 28, 2021 | |
| How to implement a layer plugin whose output shape is mutable? | 0 | 677 | November 14, 2018 | |
| Onnx to tensorrt plugin for NonMaxSuppression | 1 | 2503 | April 26, 2020 | |
| How to set max plugin limit for model with custom layers? | 3 | 311 | January 27, 2021 | |
| TensorRT Plugin: If the dynamic output shape can only be determined after the computation of the plugin, how to handle this case? | 1 | 797 | March 3, 2023 | |
| Adding a custom layer plugin for ONNX graph | 2 | 547 | August 14, 2020 | |
| ONNX Attributes aren't parsed with Custom Plugin | 3 | 1178 | December 1, 2020 | |
| Data dependent tensor shapes in TensorRT | 5 | 2535 | March 18, 2024 | |
| A qustion about custom layer plugin when i use onnx parser | 3 | 400 | April 6, 2021 | |
| TensorRT conversion from tensorflow with custom op | 5 | 1352 | August 12, 2023 |