Maybe bug in cpu backward

[pkuyym]

for (int k = 0; k < dim; ++k) { bottom_diff[i * dim + k * inner_num_ + j] *= w; } dim may should be channels?

[stormkingz]

@pkuyym I think you'are right

[mahaoyanghb]

I have checked this, I found dim=channels

[mahaoyanghb]

I have found a issue,

Backward_cpu for (int k = 0; k < dim; ++k) { bottom_diff[i * dim + k * inner_num_ + j] *= w; } should be bottom_diff[i * dim + label_value * inner_num_ + j] *=w;