All convolutions within a dense block are ReLU-activated and use batch normalization. Channel-smart concatenation is just possible if the peak and width dimensions of the information keep on being unchanged, so convolutions inside of a dense block are all of stride one. Pooling layers are insert