Giters

ujsyehao / mobilenetv3-ssd

provide pytorch model and ncnn model

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

quantize to int8 error

Alanjunhao opened this issue · comments

commented

Hi,
Thank you for your great work.
One question is that, when I use ncnn to quantize the provided .bin file to int8, Error occurred: "load_model error at layer 311, parameter file has inconsistent content." I was wondering is it because u modified .param file by hand, so there are some inconsistencies. Could you elaborate a bit 1) the reason u change the param file; 2) is there a way to quantize the bin model u provided.
Million of thanks!

commented

The reason for the error is maybe a modified param file.

  1. I modify the param file(mainly add post-process layer such as permute, flatten, priorbox layer..)
  2. I will check the quantize problem
commented

Thanks for your quick reply, looking forward to your investigation!

commented

Also, the param file generates error msg when loaded by ncnn in iOS simulator:
"
layer input not exists
layer index -1 not exists
custom layer input not exists
custom layer index -1 not exists
(lldb)
"
Hope u could also take a look at this issue, thx~

commented

@Alanjunhao Sorry, I have no IOS device.