The goal of this project is to test model against all version of tensorRT easily to assess compatiblity.
The test script is still quite basic, and requires the model to be put in the current folder.
Example :
wget https://raw.githubusercontent.com/adujardin/tensorrt-compatibility/master/check_onnx_model.sh
bash check_onnx_model.sh model.onnx --fp16
Testing TensorRT...
5.1
6.0
7.0
7.1
7.2
=====================================================
TensorRT 5.1 FAILED
TensorRT 6.0 PASSED
TensorRT 7.0 PASSED
TensorRT 7.1 PASSED
TensorRT 7.2 PASSED
For more information :
cat log_model.onnx.txt
By default the script will check against TRT 7.0 -> 8.2 ( = JetPack 4.4 and above).
The logs are displayed in real time and stored into log files, including runtimes (like "log_trt5.1_model.onnx.txt"). Since this is the stock version of trtexec, the logging isn't consistent between version and quite verbose unfortunetly. It's currently not possible to have a output like "TRT 6.0 : Compatible, took N ms (median, end to end)".
Currently supporting TensorRT versions:
4.0, 5.0, 5.1, 6.0, 7.0, 7.1, 7.2, 8.0, 8.2
The docker image used are now taken from nvidia NGC https://ngc.nvidia.com/catalog/containers/nvidia:tensorrt