All convolutions in the dense block are ReLU-activated and use batch normalization. Channel-intelligent concatenation is barely achievable if the height and width dimensions of the data remain unchanged, so convolutions in the dense block are all of stride 1. Pooling levels are inserted between dense blocks for further dimensionality https://financefeeds.com/thorchain-halts-bitcoin-and-ether-lending-amid-financial-woes/