Failed to convert my caffemodel to DLC
Join Date: 6 Dec 17
Posts: 9
Posted: Thu, 2018-05-17 00:04
Hi guys,
Now I run all examples in SNPE very well. But I got failed when trying to port my caffemodel on PC to snapdragon device.
I successfully trained this model (https://github.com/weiliu89/caffe.git (refs/heads/ssd)) on PC and tried to convert its deploy.prototxt and caffemodel into DLC file.
$SNPE_ROOT/bin/x86_64-linux-clang/snpe-caffe-to-dlc --caffe_txt $SNPE_ROOT/models/alexnet/ssd_data/deploy.prototxt --caffe_bin $SNPE_ROOT/models/alexnet/ssd_data/VGG_VOC0712_SSD_700x700_iter_32800.caffemodel --dlc ssd_data/vgg_ssd_alexnet.dlc
----error output----
/home/yangfan34/workspace/lenovo/qualcomm/snpe-1.15.0/lib/python/snpe/snpe_caffe_to_dlc.py:1567: RuntimeWarning: error_code=902; error_message=Layer parameter value is invalid in DSP. Layer conv4_3_norm: InstanceNorm (dynamic statistics calculation in BatchNorm) scale must be 1.0 on DSP runtime.; error_component=DSP Runtime; line_no=136; thread_id=139683607803712
output_name = str(self.get_output_name(layer)))
/home/yangfan34/workspace/lenovo/qualcomm/snpe-1.15.0/lib/python/snpe/snpe_caffe_to_dlc.py:2388: RuntimeWarning: error_code=1000; error_message=Layer is not supported. Layer detection_out of type SsdDetectionOutput not supported by GPU runtime; error_component=Model Validation; line_no=257; thread_id=139683607803712
confidence_threshold=params.confidence_threshold)
There is no DLC file generated in ssd_data dir. Does the error mean these two layers (conv4_3_norm and detection_out) should be removed? Hope your feedback.
Cheers,
Yangfan
Hi,
Layer detection_out is defined by SSD itself and we can skip it.
Layer conv4_3_norm seems a standard layer in Caffe. Can I know which version is supported by SNPE 1.15.0?
Cheers,
Yangfan
Hi,
I had the exact same issue. Did you manage to resolve it?
Hi,
I had the exact same issue. Did you manage to resolve it?
Not yet.
Hi yangfan34,
I have the same error with layer detection_out
how do you skip it?
Thanks