We have detected your current browser version is not the latest one. Xilinx.com uses the latest web technologies to bring you the best online experience possible. Please upgrade to a Xilinx.com supported browser:Chrome, Firefox, Internet Explorer 11, Safari. Thank you!

Showing results for 
Search instead for 
Did you mean: 
Visitor mghodsi
Registered: ‎11-16-2018

Quantizing a network using DECENT


I am trying to use DECENT to quantize resnet50 model retrieved from their official github page (https://github.com/KaimingHe/deep-residual-networks) trained for imagenet 2015. Using their caffemodel and prototxt files and following the procedure in DeePhi DNNDK v2.07 guide, it seems that the loss for calibration remains a constant value of 87.3365. It still produces the output files and then using DNNC tool I can create necessary .elf files. However the results are off when I test the network on a ZCU102 board.

Please note that I modified the prototxt input layer to ImageData type with a path to a calibration set and modified the last layers accordingly based on the examples provided in DNNDK samples. Can someone tell me if there is anything I am missing during quantization? 


Tags (3)
0 Kudos
1 Reply
Xilinx Employee
Xilinx Employee
Registered: ‎02-18-2013

Re: Quantizing a network using DECENT

@mghodsiAttached is the float prototxt of Resnet50 for your reference. It also includes the path to the calibration images from ImageNet dataset. You may need to point it to the path of your calibration images. 

0 Kudos