cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
multierection
Contributor
Contributor
608 Views
Registered: ‎05-07-2019

problem using decent to quantize yolov3.caffemodel with DNNK3.0

when I use decent to quantize yolov3.caffemodel, error reported as follows:

[libprotobuf ERROR google/protobuf/text_format.cc:274] Error parsing text-format caffe.NetParameter: 2647:20: Message type "caffe.LayerParameter" has no field named "upsample_param".
F0617 00:55:19.564172 2951 upgrade_proto.cpp:125] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: /home/yqi/Desktop/caffe/host_x86/yolov3/float.prototxt
*** Check failure stack trace: ***
./decent.sh: line 25: 2951 Aborted (core dumped) $DECENT quantize -model ${model_dir}/float.prototxt -weights ${model_dir}/float.caffemodel -output_dir ${output_dir} -method 1
root@ubuntu:/home/yqi/Desktop/caffe/host_x86/yolov3#

All the yolov3 model has this porblem. I tesed tiny/spp/mobilenet. Anyone can help me ?

0 Kudos
4 Replies
kjgreenwood
Adventurer
Adventurer
582 Views
Registered: ‎04-10-2019

0 Kudos
multierection
Contributor
Contributor
565 Views
Registered: ‎05-07-2019

Thanks for your reply, I used the the lastest version DNNDK3.0, but I got the same error as described above.

0 Kudos
kjgreenwood
Adventurer
Adventurer
558 Views
Registered: ‎04-10-2019

You need to replace your up_sample layer with something of the equivalent functionality. You can use a depthwise deconvolution layer as discussed in the link I provided (make sure you create an appropriate kernel for the type of interpolation you want). But I have also heard mention of a deephi_resize layer for doing the upsampling. Whatever you choose, you should change the name of the layer because you don't want Caffe to try to use the weights from your current yolov3.caffemodel.

0 Kudos
multierection
Contributor
Contributor
531 Views
Registered: ‎05-07-2019

Thanks,I solved by using Deephi darknet2caffe conversion tool.

0 Kudos