Hi
I trained mobilenet ssd with new dataset, I change the class_num to 195, and it can work in python,
but can't convert the network into dlc format.
Folloing is the error logs, hope someone can help me to resolve this problem. Thanks
2018-08-10 16:28:58.977108: W tensorflow/core/framework/op_kernel.cc:1318] OP_REQUIRES failed at function_ops.cc:83 : Internal: Retval[16041] has already been set.
2018-08-10 16:28:59,360 - 388 - WARNING - ERROR_TF_FALLBACK_TO_ONDEMAND_EVALUATION: Unable to resolve operation output shapes in single pass. Using on-demand evaluation!
2018-08-10 16:29:46,342 - 106 - ERROR - Conversion failed: ERROR_TF_SSD_NMS_CAN_NOT_RESOLVE_IOU: Unable to resolve IOU threshold for Multi Class Non Max Suppression layer.
I am also seeing same problem. Any update on this issue?
Same problem here!
I am not sure that is a Snapdragon SDK problem... I tried the frozen graph on linux and it's working, but when I move to windows tensorflow complain about the same thing... I don't know if this can help.
Would be nice to haveo a solution. We are a potential large scale customer but iwe need to convert our network.
Follwing updates
I am having the same issue. My command line is:
snpe-tensorflow-to-dlc --graph models/usens/200x200ssd.pb -i Preprocessor/sub 1,200,200,3 --out_node detection_classes --out_node detection_boxes --out_node detection_scores --dlc 200x200_ssd.dlc --allow_unconsumed_nodes
Where 200x200ssd.pb is a tensorflow ssd_mobilenet frozen model. I am using snpe-1.21.0. Tensorflow version is 1.12. I also tried tensorflow 1.5 which according to Qualcomm documentation is compatible with snpe1.21. The error message is as the follwoing:
[Earlier messages ignored]
2018-11-22 16:38:38,213 - 106 - ERROR - Conversion failed: ERROR_TF_SSD_NMS_CAN_NOT_RESOLVE_IOU: Unable to resolve IOU threshold for Multi Class Non Max Suppression layer.
I have tried a few other ssd models we trained all of them having this error. However, I am able to convert and run the ssd mobilenet model mentioned in the snpe tutorial.
I guess the ssd model used in qualcomm's tutorial is an older model that does not have IOU threshold parameter. But newer ssd model does support this param.
Qualcomm support please help with this issue.
Thanks,
David
I got the same issue,
u might need to downgrade tensorflow version. i guess is the nms fuction changes a lot that snpe might not adapt to it.
My experiment was using tf 1.5 and it works fine in converting.
@tohiman hi guys, thanks for you works,
the question is how to train the ssd-mobilenet in tensorflow-1.5,
i use the https://github.com/tensorflow/models/blob/master/research/object_detecti...
this course,it didn't support tensorflow<1.9,this confused me a long time,can you help me,thanks a lot