DCGAN Parameter Counting
Overview
Count how many parameters a DCGAN model needs to learn. Go layer by layer, use fixed formulas, add everything up. Pure pattern — no creativity needed.
Architecture Flow
Loading diagram...
1. Conv Layer
k² × C_in × C_out + C_out k = filter size C_in = channels IN (prev layer) C_out = channels OUT (this layer) +C_out = bias
2. BatchNorm (BN)
2 × C_out SKIP BN if: First conv (Discriminator) or Last conv (Generator)
3. GAP
ALWAYS = 0 Global Average Pooling has no learnable parameters.
4. FC Layer
C_in × C_out + C_out C_in = last conv's C_out C_out = number of classes/nodes
Exam Attack Plan
1. Read input — gray=1ch, color=3ch → that is your first C_in
2. Conv Layer 1: k²×C_in×C_out + C_out → NO BN (discriminator rule)
3. Conv Layer 2+: k²×C_in×C_out + C_out + 2×C_out (add BN)
4. GAP = 0 always
5. FC = C_in×C_out + C_out
6. Add all layers together
Common Mistakes
• BN on first conv of discriminator — always skip
• BN on last conv of generator — always skip
• GAP = some number — always 0
• Wrong C_in — must equal previous layer's C_out
• Forgetting +C_out bias in conv formula