I have build a pipeline to export a EfficientDet Pytorch implementation first to ONNX and then from ONNX to TensorRT with ONNX-Tensorrt. The conversion seems to work alright. I only get the following warning when exporting to TensorRT which shouldn’t be the cause of the problem:
Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
But when doing inference with TensorRT the output is always zero while in ONNX and in PyTorch I get a non zero output for the same input.
I also tested my TensorRT inference script with a minmal example model and it got a non zero output with the same pipeline. Am I missing something why am I only getting zero outputs?
**TensorRT Version 220.127.116.11:
**GPU Type Nvidia Jetson Xavier:
**Nvidia Driver Version N/A:
**CUDA Version 10.2:
**CUDNN Version N/A:
**Operating System + Version Ubuntu 18.04:
**Python Version (if applicable) python 3.7:
**PyTorch Version (if applicable) 1.7:
TRT model: https://drive.google.com/file/d/1kfRHI31kZfZWbfd37khhxrg8bKzmb63r/view?usp=sharing
Onnx model: https://drive.google.com/file/d/1NoawdtB2BG2Myz9dycQjjSyPm_DzyuGT/view?usp=sharing
Sample image: https://drive.google.com/file/d/1Nb_QeL96Xzrp5GsltCPTZQuCutVwiBKh/view?usp=sharing
Onnx Inference script: https://drive.google.com/file/d/1Wmgo32lb2POpFqOfJdbhYaDbs9LrgwvV/view?usp=sharing
TRT inference script: https://drive.google.com/file/d/1EFj1O5IT7ZVZQGFY0E2Vu0Qn9y51i3uN/view?usp=sharing
Steps To Reproduce
For the TensorRT conversion I used: onnx-tensorrt
Use the onnx inference script to run successful inference on the sample image
Use the TRT script to run non successful inference on the the sample image