SK-TDA4VM: segmentation fault while inferencing with TIDL delegates for TFLite SSD detection model

Hi,

Can you elaborate the issue in more depth , what is usecase you are looking specifically ?

Did model performed inference over all (5) images correctly and then after deinit seg fault occurred ?

Regards,

Pratik 

Hi,

I'm just running simple object detection with some test images from COCO dataset. The inference doesn't go through, because the script prints the results, but nothing is printed after running it with the TIDL delegates, while it works if I run it like this:

  # interpreter = tflite.Interpreter(model_path=model_path, experimental_delegates=tidl_delegate)
  interpreter = tflite.Interpreter(model_path=model_path, num_threads=2)
  interpreter.allocate_tensors()

Hi,

Could you confirm my understanding over posted thread.

1. Model Compilation was sucessful

2. Model Inference is running fine with ARM only mode ?

3. Model Inference is getting stuck with TIDL offload ?

Regards,

Pratik

Hi, 

Yes, all the 3 points are correct.

Could you able to share the model here ?

Sure, here's a WeTransfer link for the folder with the model, model artifacts and the .prototxt I used to compile the model.

And these the configs in the model_configs.py file:

    'tf2_ssd_mobilenet_v2_coco17_ptq' : {
        'model_path' : os.path.join(models_base_path,'tf2_ssd_mobilenet_v2_coco17_ptq.tflite'),
        'mean': [127.5, 127.5, 127.5],
        'scale' : [1/127.5, 1/127.5, 1/127.5],
        'num_images' : numImages,
        'num_classes': 91,
        'model_type': 'od',
        'session_name' : 'tflitert',
        'meta_layers_names_list' : os.path.join(models_base_path, 'tf2_ssd_mobilenet_v2_coco17_ptq.prototxt'),
        'meta_arch_type' : 1,
        'od_type' : 'HasDetectionPostProcLayer'
    },

The link seems to be blocked from TI's internal site.

We recommend to share all the content here as zip folder more preferably.

--

Pratik

tf2_ssd_mobilenet_v2_coco17_ptq.zip

Here it is

Thanks for sharing the model.

We will get back to you on this issue.