Coral Object detection switches from TPU to CPU

[hpmkruit]

On start i get the message that the TPU is detected and used:

16:44:59:objectdetection_coral_adapter.py: TPU detected
16:44:59:objectdetection_coral_adapter.py: Using Edge TPU

After a few seconds after the object detection module is started it switches from TPU to CPU: CPU (TF-Lite)

I tried restarting the service (host is windows) and reïnstalling the Module, but to no avail. Any tips on trouble shooting? I am using the Coral miniPCIE TPU. I tried with different models and sizes (YOLOv8 and MobileNet).

Module 'Object Detection (Coral)' 2.4.0 (ID: ObjectDetectionCoral)
Valid: True
Module Path: \modules\ObjectDetectionCoral
Module Location: Internal
AutoStart: True
Queue: objectdetection_queue
Runtime: python3.9
Runtime Location: Local
FilePath: objectdetection_coral_adapter.py
Start pause: 1 sec
Parallelism: 16
LogVerbosity:
Platforms: all
GPU Libraries: installed if available
GPU: use if supported
Accelerator:
Half Precision: enable
Environment Variables
CPAI_CORAL_MODEL_NAME = MobileNet SSD
CPAI_CORAL_MULTI_TPU = False
MODELS_DIR = \modules\ObjectDetectionCoral\assets
MODEL_SIZE = medium
Status Data: {
"inferenceDevice": null,
"inferenceLibrary": "TF-Lite",
"canUseGPU": "false",
"successfulInferences": 148,
"failedInferences": 0,
"numInferences": 148,
"averageInferenceMs": 10.675675675675675

[andmcg1]

Turn on use multi TPU. That was the issue for me. The Coral implementation is a little clunky at the moment, so that setting doesn't seem to do what you think it should and is needed

[hpmkruit]

Unfortunately this did not work. I get the error that there are no multi tpu detected and it falls back to CPU. Eventually i have gotten the model mobilenet and detlight working by just restarting the module 10x... YoloV5 and v8 never seems to work.

[whasf]

I am seeing the same issue, but the low timing on the "response received" leads me to believe that the coral TPU is still processing but the status page shows it changed to CPU.

Is there a way to verify it is actually still using the TPU or not?

[hpmkruit]

I see similar response timing (300ms on large models) but i do see an increase in CPU usage when the status page says CPU instead of TPU. Maybe check that way.