How does this skill handle memory issues during sampling?

It provides patterns for 'Detached Sampling' using stop_gradient and 'Chunked Log-Prob' to process large datasets in memory-bounded chunks without exceeding VRAM.

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

Yes, but the skill enforces that these must be wrapped in backend guards (e.g., checking keras.backend.backend()) and kept outside of the main compute path.

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

The STE pattern allows for differentiable hard indicators in calibration losses, where the forward pass uses a discrete value but the backward pass uses an identity gradient.

Why can't I use torch.sum or np.sum with this skill?

BayesFlow targets Keras 3, which supports PyTorch, JAX, and TensorFlow. Using backend-specific libraries like PyTorch or NumPy breaks compatibility for users on other backends; using keras.ops ensures it works everywhere.

BayesFlow Keras Operations

Name: BayesFlow Keras Operations
Author: matthiaskloft

bymatthiaskloft

0•

Data Science & ML

Standardizes backend-agnostic tensor math for BayesFlow extensions using the Keras 3 functional API.

This skill ensures that all tensor operations within the BayesFlow ecosystem—including custom loss functions, neural network layers, and approximator overrides—adhere to Keras 3 multi-backend standards. By enforcing the use of keras.ops instead of framework-specific libraries like PyTorch, JAX, or NumPy, it guarantees that your models and extensions remain portable across different computational backends while implementing advanced patterns like Straight-Through Estimators (STE) and memory-efficient chunked log-probability calculations.

Key Features

01Enforces keras.ops for backend-agnostic tensor mathematics

02Prevents direct dependencies on PyTorch, JAX, or NumPy in compute paths

030 GitHub stars

04Implements memory-saving patterns like detached sampling and chunked processing

05Automates proper casting and shape handling for multi-backend compatibility

06Provides standardized Straight-Through Estimator (STE) implementation logic

Use Cases

01Writing portable loss functions for simulation-based inference (SBI)

02Developing custom Keras layers for Bayesian inference networks

03Optimizing memory usage during large-scale posterior sampling

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

01Enforces keras.ops for backend-agnostic tensor mathematics

02Prevents direct dependencies on PyTorch, JAX, or NumPy in compute paths

030 GitHub stars

04Implements memory-saving patterns like detached sampling and chunked processing

05Automates proper casting and shape handling for multi-backend compatibility

06Provides standardized Straight-Through Estimator (STE) implementation logic

Use Cases

01Writing portable loss functions for simulation-based inference (SBI)

02Developing custom Keras layers for Bayesian inference networks

03Optimizing memory usage during large-scale posterior sampling

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