The module structure owns reusable parameter leaves.
The forward pass creates a new temporary computation graph.
Predict what fails if weights are created inside forward.
Value -> parameter; Module -> child; nested lists -> recursive indexed names.
Initialize registries with object.__setattr__, then intercept assignments.
This lab constructs the structure once; later list mutation is outside the registry contract.
x = [3,2]
W = [[2,-1],[0.5,1]]
b = [1,-2]
Compute both outputs and the input gradient of their sum.
Outputs are [5,1.5].
For their sum, dL/dx=[2.5,0]; each weight row receives [3,2], each bias receives 1.
Weight shape is (out,in); output stays a list even when out=1.
Linear(2,3) -> Tanh -> Linear(3,2)
First layer: 9 scalars. Activation: 0. Last layer: 8.
Expected total: 17 unique parameters. Without the activation, the composition is still affine.
The optimizer sees 34 entries but only 17 distinct identities.
Diagnosis: aliased children were traversed twice.
Deduplicate by identity, not numeric equality. Give aliases one canonical named path.
eval() recursively changes flags. These layers have no mode-dependent behavior.
It does not freeze parameters, clear gradients, or disable autograd.
The optimizer controls updates; graph construction controls differentiation.
Implement Module before layers; preserve parameter identity through every forward call.
Check the 17-parameter count, numeric affine trace, and nested mode propagation.
Zero registered gradients with correct outputs suggest a detached .data copy.