Can we save GIE model in TensorRT?

Robotics & Edge Computing Jetson & Embedded Systems Jetson TX2
1

As title, can we save GIE model?

Instead of convert caffe model to gie model every time.

Thanks!!

2

Hi r02525028, yes you can serialize the TensorRT CUDA engine after it has been imported on Jetson TX2, thus saving time when loading it in the future. See the ICudaEngine::serialize() function. You can see a code sample of it being used here:

[url]jetson-inference/tensorNet.cpp at e12e6e64365fed83e255800382e593bf7e1b1b1a · dusty-nv/jetson-inference · GitHub

3

Hi dusty,

however, can save gie model into “portable file”??

like in caffe we can save it to “.caffemodel”

in tensorflow we can save it to “.pb”

4

It is portable between systems with the same type of GPU, for example you can copy GIE model between Jetson TX2’s. Since TensorRT performs GPU-specific profiling and tuning optimizations, you will want to re-run it for different GPUs (for example, if moving to Jetson TX1 you would want to re-run).