voxl-tflite-server: "FATAL: Unsupported model provided!!"

  • I am trying to create a custom version of the voxl-tflite-server using a tflite file of the literature version of the yolo-v4-tiny model. However, when I change the code as well as the the voxl-tflite-server.conf file to include my new model I get the error "FATAL: Unsupported model provided!!" when trying to run the server. I can't find where this message is printed in the code. How can I turn this error off?

  • Dev Team

    This error is printed at line 78 of threads.cpp. If you would like to run your own model, it is much simpler to just change the




    parameters in models.cpp to match the absolute paths on voxl where these files are located (typically /usr/bin/dnn/ or in the /data/ partition if the file is larger).

  • @Matt-Turi thanks Matt. I ended up getting a segfault with it. I believe it was too memory intensive because it wasn't supported by the GPU. Is there a chance that tiny-yolo could be supported by ModalAI anytime soon?

  • Dev Team

    We've done some testing with darknet and a few tflite converted yolo models, but saw extremely poor inference times and performance from both. Is there any reason you are looking to use the yolo architecture specifically over mobilenet or some other native tensorflow model?

  • I have some yolo weights on some custom data that I collected for yolo that I wanted to test with voxl-tflite-server. Additionally, I noticed that in there are some classes missing in the coco_labels.txt file included with the code for voxl-tflite-server in master. Some classes such as 'desk' are replaced with question marks:

    0  person
    1  bicycle
    2  car
    3  motorcycle
    4  airplane
    5  bus
    6  train
    7  truck
    8  boat
    9  traffic light
    10  fire hydrant
    11  ???-11
    12  stop sign
    13  parking meter
    14  bench
    15  bird
    16  cat
    17  dog
    18  horse
    19  sheep
    20  cow
    21  elephant
    22  bear
    23  zebra
    24  giraffe
    25  ???
    26  backpack
    27  umbrella
    28  ???-28
    29  ???-29
    30  handbag
    31  tie
    32  suitcase
    33  frisbee
    34  skis
    35  snowboard
    36  sports ball
    37  kite
    38  baseball bat
    39  baseball glove
    40  skateboard
    41  surfboard
    42  tennis racket
    43  bottle
    44  ???-44
    45  wine glass
    46  cup
    47  fork
    48  knife
    49  spoon
    50  bowl
    51  banana
    52  apple
    53  sandwich
    54  orange
    55  broccoli
    56  carrot
    57  hot dog
    58  pizza
    59  donut
    60  cake
    61  chair
    62  couch
    63  potted plant
    64  bed
    65  ???-65
    66  dining table
    67  ???-67
    68  ???-68
    69  toilet
    70  ???-70
    71  tv
    72  laptop
    73  mouse
    74  remote
    75  keyboard
    76  cell phone
    77  microwave
    78  oven
    79  toaster
    80  sink
    81  refrigerator
    82  ???-82
    83  book
    84  clock
    85  vase
    86  scissors
    87  teddy bear
    88  hair drier
    89  toothbrush

  • @Matt-Turi did you try the full fledged version of yolo or was it tiny-yolo?

  • Dev Team

    @colerose Looks like that labels file is a bit outdated - I will see to updating it.

    For your other question - in the past, I tried with regular and tiny yolo implementations. With a modified darknet framework (opencl backend instead of cuda), the inference times for tiny-yolo were around 30 seconds per image, and full yolo was upwards of a minute per image. I also tested with the tflite-converted models, but saw similar performance results as well as various inconsistencies / unsupported ops due to the conversion (likely why you are getting segfaults with your model).

    For reference, any of the object detection models from the tf1 zoo as well as a few from the tf2 zoo will integrate seamlessly and can support our gpu acceleration as seen with the mobilenetv2 model that is default.

  • @Matt-Turi this is great to know, thanks Matt! 🙂 Just curious, any reason why mobilenetv2 was chosen from the tf1/tf2 zoo when it seems that there are faster and more accurate models available?

  • Dev Team

    @colerose we selected mobilenetv2 because of its exceptional performance on embedded devices - currently, inference time in the voxl-tflite-server is ~22ms per frame with very high precision. Also, the mobilenet family is fairly easy to retrain with a custom dataset!

Log in to reply