I am working on a C++ program which needs to run neural net inferences on images. My program links to TensorRT 220.127.116.11 under Ubuntu 18.04.
My models are object detection models trained with caffe and tensorflow.
I use the Faster RCNN and SSD algorithms with various feature extractors (vgg, inception, mobilenet).
I managed to convert and run the following nets
Faster RCNN - VGG16 trained with caffe
Faster RCNN - googlenet trained with caffe
SSD - inception v2 trained with tensorflow (only the pretrained model from SampleUffSSD)
but I am not able to convert and run the following nets
SSD - any arch different than inception v2 of the sample trained with tensorflow
Faster RCNN - any arch trained with tensorflow
So my questions are:
1 - is there somewhere a clear tutorial explaining how to convert and run nets for object detection trained with tensorflow (other than the sample) ?
2 - where could I find a clear explanation of how to modify the config.py config file used in SampleUffSSD for the other nets ?
If I want to use TF-TRT, I need to compile tensorflow from C sources. However, it seems from the TRT documentation that TF-TRT is not available for the C API.