Xavier NX TRT Deserialize fails for a specific caffe model

Serialization and Deserialization for
res10_300x300_ssd_iter_140000_fp16.caffemodel
works on Nano Jetpack 4.3
but fails for
Xavier NX Jetpack 4.4

Please clone this simple program that reproduces the problem


The exact same program works on Nano Jetpack 4.3
but give an assertion and segmentation fault on Xavier NX Jetpack 4.4

Hi,

Please noticed that TensorRT engine file is not portable for either library version or platform.
Do you generate the engine file directly on the XavierNX and with JetPack4.4?

Thanks.

The code both serializes and deserializes on the machine.

The code is self contained. Just clone, make, Serialize and Deserialize.
See readme.txt.

Yes. It’s done in the code in the github repo I referenced.

Hi,

Could you try to upgrade into JetPack4.4 GA which is released this week first.
There are some serializer-related fix in the new TensorRT 7.1.3 package.

More, would you mind to test the serializer-related issue with our trtexec also?

$ /usr/src/tensorrt/bin/trtexec --deploy=deploy.prototxt --output=[output/name] --saveEngine=deploy.engine
$ /usr/src/tensorrt/bin/trtexec --loadEngine=deploy.engine --verbose

Thanks.

1 Like