1x1 conv (32 x 32 x 24)  <-- initial convolution: 2 k output feature-maps

1x1 conv (32 x 32 x 48)  <-- 4 k output feature-maps.
3x3 conv (32 x 32 x 12)  <-- reducing to growth factor k
concatenate (32 x 32 x (24+12))
1x1 conv (32 x 32 x 48)
3x3 conv (32 x 32 x 12)
concatenate (32 x 32 x (24+12+12))
... (total of L=100 bottleneck-blocks)
concatenate (32 x 32 x (24+100*12))

transition layer (conv and pooling)

1x1 conv (16 x 16 x (24+100*12) / 2)  <-- from the compression, we half the number of channels
3x3 conv (16 x 16 x 12)
concatenate (16 x 16 x ((24+100*12) / 2 + 12)
...