Forums - Atrous Convolutions - DeepLabv3

2 posts / 0 new
or Register
Last post
Atrous Convolutions - DeepLabv3
calinm.cosma
Join Date: 27 Dec 17
Posts: 6
Posted: Wed, 2018-05-16 06:01
Top

Receiving the following warning while converting DeepLabv3 Mobilenet_v2:

snpe-1.15.0/lib/python/converters/tensorflow/layers/eltwise.py:145: RuntimeWarning: error_code=1004; error_message=Layer parameters combination is invalid. Layer MobilenetV2/expanded_conv_14/add: dimensions mismatch between MobilenetV2/expanded_conv_13/project/BatchNorm/FusedBatchNorm:0 {8, 8, 80} and MobilenetV2/expanded_conv_14/add {4, 4, 80}; error_component=Model Validation; line_no=117; thread_id=140575041320704
  output_name)

The network was trained with output stride 16, so when it reaches expanded_conv_14, it does not need another downsample, thus it sets stride to 1 and atrous rate to 2, in order to keep the current resolution. This should result in a 8x8 output, but when running "snpe-dlc-info" on the converted model, the output of expanded_conv_14 is 4x4.

So, are Atrous Convolutions supported by SNPE library? I do not see any information in the "Limitations and Issues" section from the Refference Guide.

 

Best regards,

   Calin

  • Up0
  • Down0
Re: Atrous Convolutions - DeepLabv3
calinm.cosma
Join Date: 27 Dec 17
Posts: 6
Posted: Wed, 2018-05-16 08:07
Top

Or does the problem come from SpaceToBatch and BatchToSpace operations?

  • 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.