ML Concepts

Conv2d Learnable Parameters

params = (kernel_h × kernel_w × C_in × C_out) + C_out
          └─────────────── weights ───────────────┘   └─ biases ┘

Example: 3×3 conv, 32 input channels, 64 output channels:

(3 × 3 × 32 × 64) + 64 = 18,496

Key points:

• Each output filter is a (kernel_h × kernel_w × C_in) 3D tensor
• There are C_out such filters
• Parameter count is independent of input spatial dimensions (H, W)
• Bias can be disabled with bias=False

Sigmoid Function

Squashes any real number to (0, 1):

σ(x) = 1 / (1 + e⁻ˣ)

• Output range: (0, 1)
• σ(0) = 0.5
• Saturates near 0 and 1 at extremes
• Used for binary classification output, gates in LSTMs

Natural Log Function

log(x), defined only for x > 0

• log(1) = 0
• log(x) → -∞ as x → 0⁺
• Grows slowly (logarithmically)
• Used in NLL/cross-entropy loss — requires clipping probabilities to avoid log(0)