mobilenet v1 inference

Hello, I am trying to get MobileNet V1 inference results in fp16 on the Jetson Nano following the instructions here:

It seems like Mobilenet-V1 224-224 input is a valid download option, and indeed I downloaded it and the script runs fine. However, on the Jetson Nano, it gives an out of memory error, and indicates that the TensorRT optimization has failed. The batch 1 inference time is 0.15 seconds, which equates to 7FPS, suggesting quite poor performance.

Did anybody run into similar problems? What should I expect to be the FPS for batch 1 fp16 inference for Mobilenet V1 with 224-224 input? I am just running image classification, not detection.

Thank you for help!


Here is our benchmark report for Jetson Nano:

The source you shared is for the detection model.
Please check this sample for the classification model:


The benchmark doesn’t seem to have results for MobileNet-V1 for image classification.

I believe that the source I shared also has a classification part. In fact, I am using stuff in the examples/classification folder of that repo.


Sorry that we don’t have exact benchmark score for MobileNet-V1 model.
But MobileNet-V1 classification model, which reaches 64FPS on Nano, should be able to give you some information.

The two tutorial is slightly different, one for TF-TRT and the other for pure TensorRT.
Since classification model is simpler, it’s recommended to use pure TensorRT for better performance.