Prerequisites & Notation

Before You Begin

This chapter builds directly on Chapter 26 (nn.Module, training loop). Familiarity with 2D convolutions from signal processing (Chapter 7) is helpful but not required.

  • nn.Module, training loop, loss functions (Chapter 26)(Review ch26)

    Self-check: Can you write a training loop with forward/backward/step?

  • Signal processing: convolution, filtering (Chapter 7)(Review ch07)

    Self-check: Do you know that convolution in the spatial domain = multiplication in frequency?

  • Image representation as tensors: (B, C, H, W)

    Self-check: Do you know that a batch of RGB images has shape (B, 3, H, W)?

Notation for This Chapter

SymbolMeaningIntroduced
mathbfK\\mathbf{K}Convolution kernel (filter) tensors01
Ctextin,CtextoutC_{\\text{in}}, C_{\\text{out}}Number of input and output channelss01
ktimeskk \\times kSpatial kernel sizes01
s,ps, pStride and paddings01
textBN(cdot)\\text{BN}(\\cdot)Batch normalisations01
oplus\\oplusElement-wise addition (skip connection)s02
← Ch 26Conv2d, BatchNorm, Activation