Open cannonli7 opened 1 year ago
That depends on the architecture where DAUs were applied. For ResNet, the DAUs replaced only 3x3 conv2d blocks inside residual blocks, while 1x1 networks remained the same. Also, for the first layer we still used conv2d with large kernel (7x7).
You can see the exact architecture we used in Figure 2 in our IJCV paper .
Thank you
Thanks for your work. I have a question about the experiment. When testing the dau, did you replace all the convolution module by dau?