I’ve tried using TensorRT with caffe models with self-define plugins, it’s relatively intuitive as we can just check the prototxt file and write the pluginFactory according to it.
But now I want to use tensorflow trained models and I have a hard time finding out which layer needs to be implemented as plugin.
For example I tied checking out the model of fasterRCNN-resnet101 on TensorBoard and the nodes layout is just too complex.
The sampleUffSSD sample helps, but for larger models it’s still a bit overwhelming.
I thought of using tools to convert frozen graph to caffe model but not sure if it’s safe to do that, and not all layers are supported by the conversion tools out there. Or maybe converts to Onnx first?
Anyone can share experience of using TensorRT with complex Tensorflow models? Thanks!