---

name: layers-and-nets description: "Use Sonnet built-in layers, normalization modules, initializers, metrics, and predefined nets with correct shapes and state." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Sonnet Layers and Nets

Use this sub-skill for Sonnet built-ins: Linear, Bias, convolutions, transpose/depthwise convolutions, embeddings, reshape/flatten, dropout, normalization, initializers, regularizers, metrics, moving averages, and snt.nets models such as MLP, Cifar10ConvNet, ResNet, and VQ-VAE quantizers.

Start here

1. Read references/api-reference.md for grouped signatures and shape/state contracts.
2. Read references/workflows.md for MLP, Conv2D+BatchNorm, ResNet, VQ-VAE, metrics, and regularizer recipes.
3. Read references/troubleshooting.md for unknown dimensions, bias conflicts, normalization state, VQ-VAE errors, and dropout flag failures.
4. Run scripts/layers_and_nets_smoke.py for a safe no-download shape/state check.

Boundaries

• Custom module design: ../module-authoring/SKILL.md.
• Training/optimizer loops: ../training-and-optimization/SKILL.md.
• Recurrent modules: ../sequence-and-rnn/SKILL.md.
• Serialization/distribution: ../serialization-and-distribution/SKILL.md.

Validation checklist

Known feature/channel dimensions are required on first call. with_bias=False forbids b_init. BatchNorm calls must pass is_training. VQ-VAE inputs must end with embedding_dim. snt.nets.MLP requires is_training only when dropout is enabled.