---

name: optimization-building-blocks description: "Routes advanced keras-vis customization requests that combine custom losses, regularizers, Optimizer control, input or gradient modifiers, backprop modifiers, and callbacks." disable-model-invocation: true metadata: disco-role: operating license: MIT

optimization-building-blocks

Use this sub-skill for low-level optimization pieces that sit underneath the higher-level activation and saliency flows.

Route away first

• For high-level activation maximization requests, use activation-maximization.
• For saliency maps or Grad-CAM, use saliency-and-cam.
• For image loading, drawing, stitching, and display helpers, use image-utilities.

Stay here for

• Writing custom Loss.build_loss() implementations.
• Combining weighted loss tuples with Optimizer.
• Choosing wrt_tensor, grad_modifier, input_modifiers, and callbacks.
• Using ActivationMaximization, TotalVariation, LPNorm, Jitter, Print, GifGenerator, and gradient/backprop modifiers directly.
• Debugging TensorFlow guided/rectified backprop overrides, graph mode, and optional callback dependencies.

Reference map

• API signatures and constraints: references/api-reference.md
• Practical workflow and composition rules: references/customization-guide.md
• Known failures and fixes: references/troubleshooting.md
• Best-effort deterministic smoke check: scripts/optimizer_smoke.py

Guardrails

• Optimizer.minimize() uses a fixed RMSProp-style loop with internal loss weighting; it only updates the input tensor itself.
• When wrt_tensor is not the input tensor, treat the loop as a gradient probe and read grads / wrt_value.
• verbose=True already adds the bundled Print callback.
• TensorFlow backprop modifiers require graph mode and do not support advanced activations here.
• GifGenerator needs both imageio and Pillow at runtime.