Forums - Faster-RCNN support

7 posts / 0 new
or Register
Last post
Faster-RCNN support
shayzweig
Join Date: 7 Feb 17
Posts: 1
Posted: Sat, 2017-10-28 13:33
Top

Hello,

Version 1.6.0 supports faster-RCNN however I am having dificulties using the model and I was wondering if you can help me.

I used the model from:

https://github.com/rbgirshick/py-faster-rcnn/tree/master/models

and  managed to convert the network to dlc from caffe but I cannot do either of the following: quantize the model or load the model as part of an android app.

When trying to quantize I need to feed in the input to the network in the input-list file but since Faster RCNN expects more than one image input I can't seem to find any documantation explaining the right format for this input file.

When trying to load the model in the android app I get an exception.  

I feel like I am missing something, however I cannot find the right reference in the documentation.

Do you have any documentation/example code showing how to work with faster-rcnn that you can share with us? 
Is there a specific pre-trained module you are supporting that I am not aware of?
Do you have any benchmark for the runtime of faster-rcnn on the DSP? 
Since this model is formally supported I believe that this knowledge exists somewhere in Qualcomm but I cannot seem to reach it.

 

  • Up0
  • Down0
Re: Faster-RCNN support
mpeniak
Profile picture
Join Date: 28 Apr 17
Location: London
Posts: 6
Posted: Tue, 2017-10-31 05:27
Top

I am equally interested in this...have you made any progress?

Thanks,
Martin

  • Up0
  • Down0
Re: Faster-RCNN support
jesliger
Join Date: 6 Aug 13
Posts: 75
Posted: Wed, 2017-11-01 10:41
Top

SNPE 1.6.0 snpe-dlc-quantize cannot quantize models with multiple inputs or outputs.  So this model cannot be quantized offline using the snpe-dlc-quantize tool.  But, the SNPE runtime will quantize it on-the-fly when it is deploying it to the DSP. 

FasterRCNN support in SNPE has some limitations and some slight deviations from the public py-faster-rcnn project.  There is a bit of documentation in the user's guide, limitations section (on Proposal layer and ROIPool).

We will try to provide further documentation on this.  

  • Up0
  • Down0
Re: Faster-RCNN support
huawan0
Join Date: 28 Dec 17
Posts: 3
Posted: Thu, 2017-12-28 19:11
Top

How can I set the value of PRE_NMS_TOP_N and POST_NMS_TOP_N for the Proposal Layer ?

  • Up0
  • Down0
Re: Faster-RCNN support
zf.africa
Join Date: 15 Jun 17
Posts: 51
Posted: Thu, 2018-01-04 06:57
Top

Hi jesliger,

I tried to set multiple output layers when load container, but I got NULL pointer for snpe, and I cannot get the error log about why snpe is set to NULL. Is it a bug of SNPE SDK?

Thanks!

  • Up0
  • Down0
Re: Faster-RCNN support
huawan0
Join Date: 28 Dec 17
Posts: 3
Posted: Mon, 2018-01-22 02:15
Top

Dear engineers at Qualcomm,

Is there any plan to write code for the Proposal Layer that can run on GPU ?

 

 

  • Up0
  • Down0
Re: Faster-RCNN support
zakai
Join Date: 24 Jul 18
Posts: 3
Posted: Wed, 2018-07-25 04:10
Top

hi jesliger, 

  According to the limitations of Proposal layer and ROIPool, does  "max_num_rois" (an attribute of layer proposal) can  only be set to "1",  and other options doesn't make any sense? my snpe version is 1.17.0.

 Looking forward to your reply, thanks!

  • Up0
  • Down0
or Register

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.