Custom segmentation model on TX2

Hello everyone. I have a problem with a custom model (I meant custom architecture). It was assembled with Keras and then converted to .onnx format. I was guided by the git tutorial(@dusty_nv) and I have problems with paddings inside the model. When I surfed on the Internet’s waves I found an article which explains that it’s impossible to work with paddings in condition tensorflow-onnx->tensorrt. Link with explanations, second link with more details about paddings in TRT. I just want to ask moderators how to solve this issue without changing networks’ architecture. Maybe It launched incorrectly or maybe I missed somewhere a specific instruction. Anyway, I need you advice, please.


This is not supported. The padding need to be applied along the two innermost dimensions:


Thank you) Kind regards.