Tensorflow Object Detection - Improve loading time?

I have downloaded SSD ResNet50 V1 and converted it to a tflite model since I am going to use it on a Jetson Nano. The loading time is about 226 seconds.
If I do the same operation with MobileNet v2 it takes about 195 seconds to load the model

Of course, for an application to use over 3 minutes to get ready is a little bit high in my opinion. If for example the application crashes and has to be restarted it can already be to late doing predictions as the object might be gone.
Is there anyway I can make the loading time improve?

Hi @TensorOverflow ,

I got a response from a TensorFlow Advocate:
" Jetson Nano uses TF for inference, not TFLite nor TFLite Micro.

TF has a benchmark tool: GitHub - tensorflow/benchmarks: A benchmark framework for Tensorflow

As I’m not familiar with Jetson Nano, I’m not sure if this tool can run on the device but it’s worth a try.

If one wants to do object detection on an IoT device, I’d advise them to use Coral. AFAIK, Coral doesn’t have this slow initialization issue."

Have you tried to use TF with Tensorrt on Jetson Nano?


Sample Support Guide :: NVIDIA Deep Learning TensorRT Documentation