cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
guo_bing
Adventurer
Adventurer
297 Views
Registered: ‎05-27-2019

Different weights on the 1st layer between DNNDK3.0 and DNNDK2.08 with the same model, is it reasonable?

The same caffe model, provied by DNNDK in the dnndk/host_x86/model/caffe/resnet, should it have the same weights file dumped in dexplorer -m debug mode? 

1. The example resnet50 model -> DNNDK v2.08 decent.sh, dnnc.sh -> build into application -> run on DPU1.3.0, ARCH: B1152  (From DPU Integration Tutorial)

2. The example resnet50 model -> DNNDK v3.00 decent.sh, dnnc.sh -> build into application -> run on DPU1.4.0, ARCH: B2304   (From prebuild image: xilinx-ultra96-prod-dpu1.4-desktop-buster-2019-05-31.img.zip)

I enable debug mode by dexplorer -m, and dump all layer's information, compare them, I found that the first layer's weight file are completely different, but almost the same for the rest layers.

Can I have a basic assumption that a model should have the same behavior even under different DPU versions, architecutures and DNNDK tools? The reason I dump all the layer's log is that I found the output of the first layer is nearly all zero, so I dump them to look what happen inside.(I had post this phenomenon here.)

resnet50_0_conv1_weights.bin should be the 1st layer's weights:

Image_20190704-202657.png

The details of the resnet50_0_conv1_weights.bin:  Completely different ! really confused.

Image_20190704-203043.png

0 Kudos
0 Replies