Why is the Coral USB not running when executing own tflite model?

Question

Setup: Raspberry PI 4b 4GB with Coral USB accelerator.

I am running this detection program which is from the official Coral USB website.

Running detection with mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite works at 30FPS. And the LED on the Coral USB pulses, which means the Edge TPU is running .

python3 detect.py --model=../all_models/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite --labels=../all_models/coco_labels.txt

And the output is a stream of something like this:

96% person Inference: 11.99ms FPS:22.9

Running mymodel.tflite runs at 3FPS. The LED on the coral USB is solid, which means the Coral USB is initialized but not running.

python3 detect.py --model=../all_models/mymodel.tflite --labels=../all_models/dict.txt

And the output is a stream of something like tis:

85% S_13 Inference: 290.25ms FPS:3.2

The examples with gstreamer, pygame and raspicam in the GitHub repo yield the same result (3FPS) and solid light if I run my own model but all run at 30FPS with mobilenet and the LED blinks in all three cases.

I believe the problem has to do with the Coral USB not running when the Raspberry executes my own model, which has these specs

Other stuff

My labels are in the same format as coco_labels.txt and the detection, even if slow, happens. Thus, the labels do not seem to be the problem.

Answer 1

Sounds to me that you haven't pas your model through the edgetpu_compiler .

edgetpu_compiler -s mymodel.tflite

If it fails, check if your custom model passes the requirements here: https://coral.ai/docs/edgetpu/models-intro/#model-requirements