---

name: automodel-customization description: "Build custom AutoKeras AutoModel graphs, blocks, multimodal inputs, and multitask outputs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

AutoModel customization

Use this sub-skill when a task needs more control than the one-class task APIs provide: custom blocks, functional graph topology, multimodal inputs, multitask outputs, or explicit node/block/head wiring.

Fast route

1. Choose input nodes: ak.ImageInput(), ak.TextInput(), ak.StructuredDataInput(...), or generic ak.Input().
2. Connect nodes through blocks and reductions using Keras-functional style calls.
3. End every output branch with a ClassificationHead or RegressionHead.
4. Build ak.AutoModel(inputs=..., outputs=..., max_trials=..., overwrite=...).
5. Pass x and y to fit in the same nesting/order as inputs and outputs.

Read references/api-reference.md for verified AutoModel, node, block, and head signatures. Read references/custom-search-spaces.md for custom topology recipes. Read references/multimodal-multitask.md for ordered multiple input/output workflows. Read references/troubleshooting.md for graph connectivity, array-count, backend, pretrained, and runtime failures.

Bundled helpers

Run helpers with --help first. They default to safe dry-run construction and use synthetic data.

• scripts/build_tiny_custom_image_automodel.py builds a small image custom search graph.
• scripts/build_tiny_multimodal_automodel.py builds a two-input, two-output multimodal/multitask graph.

Route elsewhere

• High-level ImageClassifier, TextClassifier, or StructuredDataClassifier style workflows: ../task-apis/SKILL.md.
• Tuner selection, search directories, export/reload, callbacks, objectives, or old search cleanup: ../search-and-export/SKILL.md.