LeNetConfig

Configuration for LeNet-5.

activation controls the nonlinearity:

• "tanh" — plain $\tanh$. The 1998 paper's squashing function is the scaled $f(a) = 1.7159\tanh(\tfrac{2}{3}a)$ (Appendix A), chosen so $f(\pm 1) = \pm 1$; plain tanh is what every modern re-implementation uses and what the shipped weights were trained with.
• "relu" — modern convention

pooling controls the sub-sampling layers:

• "avg" — original paper (average pooling / sub-sampling)
• "max" — modern convention

in_channels defaults to 1 (grayscale). Set to 3 for RGB inputs, though the canonical use-case is MNIST / single-channel images.