วางชั้นเชิงเส้นซ้อนกัน

A single nn.Linear layer maps inputs to outputs with a weighted sum plus a bias. It reshapes data from one size to another.

Your first layer takes the raw input features and projects them into a hidden space, often a larger or smaller width.

A later layer maps the hidden features down to your final output size, like the number of classes you want to predict.

Each layer's out_features must equal the next layer's in_features. These matching dimensions form a clean pipeline.

Stacking bare linear layers stays linear, so place an activation like ReLU between them to unlock real depth.

More neurons per layer means more width; more layers means more depth. Both add capacity in different ways.

Picking the hidden size is a design call. Bigger learns more patterns but costs memory and risks overfitting.

In forward, the input passes through layer one, then the activation, then layer two. Each step transforms the tensor in turn.

Input, hidden, output with a nonlinearity is the classic multilayer perceptron. It is the workhorse of feedforward nets.

Every linear layer adds its own weights and bias. Stacking more layers steadily grows the model's parameter count.

Begin with one hidden layer and expand only if needed. A lean stack trains fast and is easy to debug. 🌱

Recall what must sit between two linear layers to make stacking worthwhile.

Stack nn.Linear layers so each output size feeds the next input size, with an activation between them. That builds a real MLP. 🎯