detection performance degrade seriously after Decent quantization

[zfzfdafei]

Hi everyone.I have encounter a problem when I use resnet50 as SSD backbone. When I trained the network after 20000 iterations in my own dataset(grayscale,11 classes) the detection eval reach 92%, but when I use Decent quantization, after calibration, the test detect_eval is only 8%. The carlibration dataset have 135 images.

I have use resnet50 as SSD backbone to train VOC dataset before, no such degradation happened in the process of Decent quantization, only +/- 5% around the caffe test result.

I have tried the decent weight_bit data_bit method and calib_iter parameters,but the result is still poor, can anyone tell me the reason?

[quentonh]

@zfzfdafeiA few suggestions...

[1] Please review jcory's commentary related to batchnorm / scale layers at this post:

https://forums.xilinx.com/t5/Deephi-DNNDK/An-error-about-DNNC/m-p/995277

[2] Please ensure that you are using images from your training set for calibration

[3] If possible, please increase the number of calibration images that you are using to 100 images

--Quenton