What is the Straight-Through Estimator (STE) pattern used for?

The STE pattern provided by this skill allows for differentiable hard indicators, which are essential for specific types of calibration loss in probabilistic modeling.

Why should I use keras.ops instead of direct PyTorch or NumPy calls?

Using keras.ops ensures your code is backend-agnostic, allowing it to run on PyTorch, JAX, or TensorFlow without any code changes, which is a requirement for the BayesFlow ecosystem.

How does this skill handle memory optimization?

It provides patterns for 'Detached Sampling' to save activation memory and 'Chunked Log-Prob' to process large datasets in memory-bounded intervals.

Can I still use NumPy arrays with this skill?

Yes, but only at the boundaries. The skill guides you to use keras.ops.convert_to_tensor to bring NumPy data into the Keras compute path safely.

Keras 3 Multi-Backend Operations

Name: Keras 3 Multi-Backend Operations
Author: matthiaskloft

bymatthiaskloft

0•

Data Science & ML

Enforces backend-agnostic tensor math using Keras 3 to ensure code compatibility across PyTorch, JAX, and TensorFlow.

The keras-ops skill provides specialized guidance for writing portable tensor mathematics within the BayesFlow ecosystem and Keras 3 projects. It ensures that developers utilize the keras.ops API instead of backend-specific libraries like PyTorch or JAX, preventing common compatibility errors in multi-backend environments. By implementing best practices for loss functions, custom layers, and network forward passes, this skill helps maintain a single codebase that runs seamlessly across different computational frameworks while providing optimized patterns for detached sampling and memory management.

Key Features

01Standardizes tensor operations using the backend-agnostic keras.ops API

02Provides mapping references for common PyTorch and NumPy operations to Keras

03Enforces memory-efficient patterns like detached sampling and chunked log-probs

04Implements Straight-Through Estimator (STE) patterns for differentiable logic

05Detects and corrects boundary-crossing errors between NumPy and Keras tensors

060 GitHub stars

Use Cases

01Developing custom BayesFlow extensions and loss functions

02Creating portable Keras layers compatible with PyTorch, JAX, and TensorFlow

03Optimizing memory usage for large-scale simulation-based inference models

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add matthiaskloft/claude-skills keras-ops

For use in Claude.ai and ChatGPT

Key Features

01Standardizes tensor operations using the backend-agnostic keras.ops API

02Provides mapping references for common PyTorch and NumPy operations to Keras

03Enforces memory-efficient patterns like detached sampling and chunked log-probs

04Implements Straight-Through Estimator (STE) patterns for differentiable logic

05Detects and corrects boundary-crossing errors between NumPy and Keras tensors

060 GitHub stars

Use Cases

01Developing custom BayesFlow extensions and loss functions

02Creating portable Keras layers compatible with PyTorch, JAX, and TensorFlow

03Optimizing memory usage for large-scale simulation-based inference models

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add matthiaskloft/claude-skills keras-ops

For use in Claude.ai and ChatGPT