Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

convolution_3x3.h 中 conv3x3s1_winograd64_neon5/conv3x3s1_winograd64_neon4 效果不对 #517

Closed
spacegrass opened this issue Aug 6, 2018 · 6 comments

Comments

@spacegrass
Copy link

我试了一下PC/ANDROID两个版本的winograd,发现效果都不对;64位版本。这个有人发现类似问题么?

@cicido
Copy link

cicido commented Aug 7, 2018

我发现另一个是pool层结果不对。
pc Mat结果:
tmp h w c: 11,11,28
-0.267136 -0.274775 -0.309783
-0.254186 -0.259913 -0.253850
-0.288264 -0.259913 -0.253850

2.042570 2.042570 1.477021
0.007943 0.166107 0.304524
0.261186 0.134180 0.544117

0.012891 0.014708 0.000303
0.078063 0.106060 0.141780
0.283883 0.287117 0.269038

android结果:
tmp h w c: 11,11,28
-0.267136 -0.274775 -0.309783
-0.254187 -0.259913 -0.253850
-0.288264 -0.259913 -0.253850

0.152344 0.160156 0.542969
0.007943 0.166107 0.304524
0.261186 0.134180 0.544117

0.012891 0.014708 0.000303
0.078063 0.106060 0.141780
0.283883 0.287117 0.269038

从源代码中打印的结果来看,pc端跑的代码与android跑的代码是一样的。android中还用到了汇编代码。

@cicido
Copy link

cicido commented Aug 7, 2018

我是逐层打印Mat结果,Conv层结果没有差别。

@nihui
Copy link
Member

nihui commented Aug 7, 2018

请提供下 conv pooling 的调用参数 kernel stride pad 等,以及输入blob的w h c

@spacegrass
Copy link
Author

bottom_blob: w: 74, h: 102, c: 40, n: 1; cstep: 7548
top_blob: w: 72, h: 100, c: 16, n: 1; cstep: 7200
cnn_kernel: w: 3, h: 3, c: 40, n: 16, cstep: 12
bias: w: 16, h: 1, c: 1, n: 1, cstep: 16

bottom_blob已经做了copy_make_border, pad_w == pad_h == 1

请帮忙分析一下,谢谢~

@nihui
Copy link
Member

nihui commented Aug 9, 2018

@spacegrass x86 和 armv7 aarch64 上测试没有问题,检查下你自己做的 padding 是不是有问题

@spacegrass
Copy link
Author

问题解决了,没有创建winograd的kernel,谢谢

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants