Why should I use keras.ops instead of torch or numpy?

Using keras.ops ensures your code is backend-agnostic, allowing it to run seamlessly on PyTorch, JAX, or TensorFlow without modification, which is essential for the BayesFlow ecosystem.

Can I still use backend-specific features like CUDA cache clearing?

Yes, the skill includes guidance on using backend-specific code guards that allow you to call features like torch.cuda.empty_cache() without breaking non-PyTorch backends.

How does this skill handle gradient stopping?

It provides patterns using keras.ops.stop_gradient to cut gradients on sampling paths, which is a critical optimization for saving activation memory during log-probability evaluations.

How do I convert NumPy arrays to Keras tensors safely?

Use the keras.ops.convert_to_tensor() function at the boundary where data enters your Keras compute path to ensure proper type handling and backend compatibility.

Keras Multi-Backend Ops

Name: Keras Multi-Backend Ops
Author: matthiaskloft

bymatthiaskloft

0•

データサイエンスとML

Standardizes tensor mathematics for BayesFlow extensions using Keras 3's backend-agnostic operations to ensure compatibility across PyTorch, JAX, and TensorFlow.

This skill provides a specialized framework for writing portable tensor math within the BayesFlow ecosystem. By enforcing the use of Keras 3 ops instead of backend-specific libraries like PyTorch, JAX, or NumPy, it ensures that custom layers, loss functions, and network forward passes remain compatible regardless of the underlying computational engine. It includes optimized patterns for detached sampling, memory-efficient chunking, and straight-through estimators, helping developers avoid common pitfalls such as backend-specific API leaks or unintended gradient flow issues.

主な機能

01Includes strict boundary-crossing guards for NumPy and Keras tensor conversions

020 GitHub stars

03Implements advanced patterns like Straight-Through Estimators (STE) and detached sampling

04Provides a comprehensive mapping of common tensor operations to their backend-agnostic equivalents

05Offers memory-efficient chunking strategies for processing large posterior sample sets

06Enforces Keras 3 keras.ops for universal backend compatibility across PyTorch, JAX, and TensorFlow

ユースケース

01Writing custom Keras layers or loss functions for BayesFlow that work across all supported backends

02Optimizing memory usage in Bayesian inference through gradient stopping and chunked log-probability evaluations

03Implementing complex neural network forward passes without hard-coding PyTorch or JAX dependencies

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

主な機能

01Includes strict boundary-crossing guards for NumPy and Keras tensor conversions

020 GitHub stars

03Implements advanced patterns like Straight-Through Estimators (STE) and detached sampling

04Provides a comprehensive mapping of common tensor operations to their backend-agnostic equivalents

05Offers memory-efficient chunking strategies for processing large posterior sample sets

06Enforces Keras 3 keras.ops for universal backend compatibility across PyTorch, JAX, and TensorFlow

ユースケース

01Writing custom Keras layers or loss functions for BayesFlow that work across all supported backends

02Optimizing memory usage in Bayesian inference through gradient stopping and chunked log-probability evaluations

03Implementing complex neural network forward passes without hard-coding PyTorch or JAX dependencies