This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

  • Resolved

AM5749: Merging reshape layer into softmax layer in caffe

Prodigy 30 points

Replies: 3

Views: 104

Part Number: AM5749

Hi,

I had a problem doing quantization for a detection network in caffe in which I would like to use a reshape layer followed by a softmax layer. I got an error code of -1100 during the quantization process using SDK tool. I referred to AM5749 as well as this link (http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components_TIDL.html) which shows that using reshape layer with softmax layer is feasible in SSD "context". So I just want to know is it possible to do it the same way in my case since it's similar as in SSD.

ti_caffe_reshape.zip

I have enclosed the caffe .prototxt and .caffemodel in case you need to do an analysis.

mnet.prototxt : structure file for network with reshape and softmax

mnet_no_reshape.prototxt: structure file for network without reshape and softmax

mnet.caffemodel: weights file for network both with and without reshape and softmax

RETINA.BIN and RETINA.PRM: output file of quantization

Would appreciate if you could give a better explanation how to use reshape followed by softmax. Thanks in advanced!

  • Sorry, I guess I have given the wrong the part number, it should be tda2xx.

  • In reply to yipinzhu:

    Hi,

    A DetectionOutput layer must be present in a network model in order for TIDL import tool to recognize it as an “SSD” network. Your model doesn’t have a DetectionOutput layer and is not recognized by TIDL import tool as SSD network. Therefore, Reshape is not merged with Softmax.

    Below is our JDetNet model which has a DetectionOutput layer and therefore Reshape layer is merged into Sofmax.

     

      Regards,

    Manisha

  • In reply to manisha:

    Hi Manisha,

    Thanks for the clarification, this answers my quesiton for the time. While I hope you can soon support stand alone Reshape layer and Softmax layer for detection networks.

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.