in the First component of the convolutional_block() function, the kernel size is 1, but the stride depends on the param s , so i got a question: when s > 1 , will Conv2D function miss some features of the input X ? meanwhile the shotcut path Conv2D function also miss the same features. The bigger the s setting, the more features will be lost, and I don’t see anything to make up for it, so I think it’s a little bit inappropriate
It is a good observation. I think you’re right: if the stride is > 1, then it will actually ignore some inputs completely. I do not have an explanation. We are just given this Residual Network architecture. The stride in the conv block function has a default value of 2 and is an optional parameter. It’s possible that in “real” applications of this they usually call this with a stride of 1, but I don’t know for sure.