Symbols used

• Layer
• Number of nodes/units
• Input width
• Input height
• Input channels
• Number of filters used
• Filter size
• Padding
• Stride

FC Layers

Output Dimensions:

Parameter counts:

CONV and POOL Layers

Output Dimensions:

where

$$
n^{[l]}C = n^{[l]}{C’}
$$

• POOL often has and
• If we want “same” padding (so input size equals output size), we have to choose

Parameter counts:

$$(f^{[l]} \times f^{[l]} \times n^{[l]}C + 1) \times n^{[l]}{C’}$$

• The is for the one bias per filter