How to implement a forward pass of identity matrix?

Hello everybody,

can someone please tell me if we can compute forward pass of an identity matrix in CNN?

For instance, if we use fully connected layers only, we can definitely feed an identity matrix to the network and get some results.

However, what if we use convolutional layers? We know convolutions are just linear operations, so there must be some equivalent fully connected layers, which can replace the convulutional layers.

Therefore, can we find an input, so that CNN is “equivalently” computing the forward pass of an identity matrix?

thank you!