Hey! I am trying to run my own yolov3 model with lstm and spp layers in the object_detection_demo_yolov3_async.py script.
My model has been successfully frozen, and also successfully standardized for OpenVINO, but when the script is run, the following happens: the model produces an infinite number of bounding boxes with an accuracy value of 1.00. The original model works well. Please, tell me how to fix this error?
Your model managed to be run with the object_detection_demo.py Python demo, however there are no bounding boxes or etc.
I believe you'll need to train this model first. You may refer to my attachment.
The trained model should give out the output as my 2nd attachment. (I used yolo-v3-tf pretrained model with the demo app).
Another thing to take note, the object detection sample app in the latest OpenVINO is named object_detection_demo.py and you will need to use the -at (ssd,yolo,faceboxes,centernet,retinaface):
I suggest that you use the latest OpenVINO version with the latest demo/samples application.
Hello again. I am having a problem when starting a custom model with an lstm layer added. I have trained the network for custom classes (stickers). I successfully launched the classic network configuration on openvino, the problem is just when adding custom layers
I am attaching an example from the test sample
May I know your reference source for LSTM attributes that you had added into your model (bottleneck & output) as from this TF link, none of these attributes are added into the LSTM layer - https://www.tensorflow.org/api_docs/python/tf/keras/layers/LSTM#attributes
Plus, did the custom model (yolov3) is implemented in TF2 Keras?
Thank you for your patience. This reply took quite some time due to the delay from the engineering team's perspective.
Currently, it is confirmed that the "bottleneck" attribute is not supported.
Fyi, you can refer to this release note for updated features once the new release is out:
Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question.