I am trying to run keras-retinanet (https://github.com/fizyr/keras-retinanet) on my Jetson Nano board. First, I installed keras-retinanet on the board itself without any hassle but when I tried to run inference, it resulted in FailedPredictionError. I did try with
TF_CUDNN_USE_AUTOTUNE=0 as an environmental variable, unfortunately, that did not work as well.
The next thing I tried was to get the weights from retinanet model in .h5 format, then convert it to its TRT variant. I am not sure if I am doing that in the right way, but I am trying to run inference using the TRT graph, it is causing FailedPredictionError.
Retinanet has been super successful in recent object detection tasks over two-stage detectors and single-shot detectors. This is why I thought, running retinanet on a Jetson Nano would be damn cool.
Here’s the link (https://www.dropbox.com/sh/5u1a1sxd6c51rl0/AADj_zcPj0X-IWM8gaTAxE7Ta?dl=0) which contains the script that converts .h5 weights to a TRT graph, the relevant error trace, and an example notebook with which I am trying to run the inference using the TRT graph.
Any help would be immensely appreciated.