kevinbro96
kevinbro96
You can try to replace line 612-627 in perceptual-advex/perceptual_advex/perceptual_attacks.py with the following: updates = torch.zeros_like(inputs) updates[live] = -grad_normed * ( step_size / (dist_grads + 1e-8) )[:, None, None, None] if...
The MXNet implementation only consider layers: Convolution, Deconvolution, FullyConnected, Pooling, relu. But the SplAtConv2d contains many split attention modules as well as rSoftmax modules which are not included in the...
Thank you, but I don't agree with you. The GFLOPS of channel-wise multipication is at least equal to the size of the output feature map, because each point in the...
I'm afraid not. I have tested the latency (time of forwarding a image) of SENet and ResNest. The latency of SENet101 is 1.2s, while that of ResNest101 is 2.2s. Although...
I find the same question.
你好，谢谢你的回答，但是#4和我问的没有什么关系。我想问为什么要先把特征x先做L2norm后再计算loss，x.norm(p=2, dim=1)，这样不是会丢失特征x的真实模值吗?而且，就算不把x先做L2norm，也可以一样达到你想要的loss只和\Delta r有关的效果。