Why should I use keras.ops instead of native PyTorch or JAX?

Using keras.ops ensures that your code is backend-agnostic, allowing it to run seamlessly on any Keras 3-supported framework without requiring rewrite for different environments.

Does this skill support backend-specific guards?

Yes, it provides patterns for using keras.backend.backend() to safely implement truly backend-specific features like CUDA cache clearing without breaking the main compute path.

How does this skill handle memory optimization?

The skill provides specific patterns for detached sampling and chunked log-probability calculations to minimize activation memory and handle large posterior sample sets.

Can I still use NumPy with this skill?

You can use NumPy for data preparation, but the skill enforces conversion to Keras tensors at the boundary using keras.ops.convert_to_tensor to maintain consistency within compute paths.

BayesFlow Keras-Ops

Name: BayesFlow Keras-Ops
Author: matthiaskloft

bymatthiaskloft

0•

데이터 과학 및 ML

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

This skill provides domain-specific guidance for writing and modifying tensor math within BayesFlow extensions using the Keras 3 ecosystem. By enforcing the use of keras.ops instead of backend-specific libraries like Torch or NumPy, it ensures that loss functions, custom layers, and network forward passes remain fully backend-agnostic. The skill includes optimized implementation patterns for detached sampling, memory-efficient chunked log-probability calculations, and straight-through estimators (STE), helping developers maintain high-performance machine learning code while avoiding common pitfalls such as gradient flow breaks and backend lock-in.

주요 기능

01Optimizes memory usage through detached sampling patterns

02Standardizes boundary conversions between NumPy arrays and Keras tensors

03Implements Straight-Through Estimators (STE) for differentiable indicators

04Enforces backend-agnostic tensor math using the keras.ops API

05Provides templates for memory-bounded chunked log-probability processing

060 GitHub stars

사용 사례

01Optimizing memory-intensive posterior sampling paths in neural networks

02Developing custom Keras layers that work across PyTorch, JAX, and TensorFlow

03Implementing backend-agnostic loss functions for BayesFlow extensions

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

주요 기능

01Optimizes memory usage through detached sampling patterns

02Standardizes boundary conversions between NumPy arrays and Keras tensors

03Implements Straight-Through Estimators (STE) for differentiable indicators

04Enforces backend-agnostic tensor math using the keras.ops API

05Provides templates for memory-bounded chunked log-probability processing

060 GitHub stars

사용 사례

01Optimizing memory-intensive posterior sampling paths in neural networks

02Developing custom Keras layers that work across PyTorch, JAX, and TensorFlow

03Implementing backend-agnostic loss functions for BayesFlow extensions