Data loss in ResNetv50

Hi Coursera community,

I have a question to made clear, that is in ResNet50, some convolutional block is initialized with 1x1 kernel size but stride > 1. As I know 1x1 convolution helps to reduce the number of filters, but stride > 1 will cause some pixels to be lost?

I hope to receive helps from you guys,


It is a good point that in the “conv block” both the main path and the shortcut path include a 1 x 1 convolution with a stride of 2, so it literally ignores half of the input pixels. I don’t know why they do this as opposed to using a pooling layer to reduce the dimension. It might be worth having a look at the original paper about Residual Networks to see if they comment on this point.

1 Like

thanks for your information, I would have a look then :wink: