Hi Qualcomm,
I am working on SNPE, and find it's very efficient for CNN. Depthwise separable convolution becames a prevalent layer in CNN such as Xception, MobileNet, but I found the SNPE does not support it. So I have to use UDL to implement it myself, but the performance of my implementation is not very good.
Would you like to add this feature to convolution layer in the future?
Thanks.
Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.
Hey yaochuanqi888,
As you have mentioned you tried UDL and succeeded. Could you please guide me how to use this UDL feature of snpe-sdk . I am trying for mobilenet model. I have a trained one (and have a frozen model). There's a snpe-caffee-to-dlc-udl option there provided with the sdk but not for tensorflow (as my mobilenet model is based on tf not caffee). So, how to proceed further?
Regards,
Lucky Srivastava
My mobilenet model is based on caffe:
https://github.com/chuanqi305/MobileNet-SSD
Hi yaochuanqi888,
Could you share yuor implementation for snpe-caffe-dl-udl code for conversion, if possible.
It would be of great help to understand the use of udl API's in SNPE.
Also could share the performance timing on GPU/DSP?
Thanks and Regards
Vishal