All convolutions in a very dense block are ReLU-activated and use batch normalization. Channel-wise concatenation is barely feasible if the peak and width dimensions of the info continue being unchanged, so convolutions inside of a dense block are all of stride one. Pooling layers are inserted b