Error converting MobilenetSsd_v2 with conversion script

Technical Support
1

i got a warning message when i converted ssd_mobilenet_v2.pb (frozen model from TensorFlow model zoo) ( 0_import.sh and 1_quantize.sh script works very well, when i use 2_convert script.sh i got a warning)
like this
W extend FeatureExtractor/MobilenetV2/expanded_conv_7/add_142 to add will cause accuracy loss, do not extend. W extend FeatureExtractor/MobilenetV2/expanded_conv_5/add_175 to add will cause accuracy loss, do not extend.

and my network worked on vim3 .always gave the wrong result.when i searched on net ,i found this same problem(they have rockchip ,and he use Rocktoolkit) http://t.rock-chips.com/forum.php?mod=viewthread&tid=747&page=1#pid3480
i think its because ACUTY TOOL,this tool dont convert same connection (rockchip toolkit uses also ACUITY) i want to share with screenshot ,in here is expanded_conv2/add problem

Ekran Görüntüsü - 2019-12-12 22-03-26
Ekran Görüntüsü - 2019-12-12 22-03-261920×1080 441 KB

and here is expanded_conv4/add .They affect my all result
Ekran Görüntüsü - 2019-12-12 22-05-3322
Ekran Görüntüsü - 2019-12-12 22-05-33221920×1080 433 KB

2

@ahmetkemal If this error comes from the transformation tool, we can’t fix it, because this tool is not developed by us

3

İ know this reason for this is acuity tool. The same tool uses rockchip-neural network tool. İs it a problem for me to contact verisilicon? You know that npu tool has license problem. Would it be a problem for khadas, if verisilicon finds out that I have such a tool?I didn’t want to contact verisilicon without consulting Khadas.
Can i contact to verisilicon??

4

@ahmetkemal We already have a license . You can ask verisilicon . There’s no problem with that.

5

I want to share my experiences,
all problem should be related to ssd_mobilenet_v2_coco architecture,
i think ACUITY and NPU conversion tool cannot convert inversion this Neural Network.
The same tools (conversion script) can convert ssd_mobilenet_v1_coco and ssd_inception_v2_coco (my retraining network )
The result works great with these two networks.,

Please pay attention to the warning messages when running the last script (2_export_case_code.sh script in folder conversion_scripts)
such a message indicates that the neural network will work incorrectly:
W extend FeatureExtractor/MobilenetV2/expanded_conv_5/add_175 to add will cause accuracy loss, do not extend

Try other neural networks ,and don’t waste time with retraining ssd_mobilenet_v2_coco,ssdlite_mobilenet_v2_coco
I would recommend ssd_mobilenet_v1_coco,ssd_inception_v2_coco for retraining(maybe you can use yolov2 and yolo3 i didnt use)

You can find these networks here. https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

And would like to share this little result with you :grinning:the screen resolution is 4k and fullHD video, you can get 15 fps in this way. https://youtu.be/oT2QlKKk2XM
Pay attention to temperature:hot_face:
I hope it will help,Thanks.

1 Like