Can this skill prevent CUDA out-of-bounds errors?

Yes, it provides implementation patterns for safe indexing in replay buffers, ensuring that 'next state' transitions never exceed the buffer's bounds, especially for the final transition in a dataset.

Does it help with numerical stability?

Absolutely. It includes mandatory NaN/Inf checks and minimum sample size validation to prevent corrupted distribution fitting or diversity metrics that can lead to training failure.

How does this skill handle LayerNorm with Conv1d?

It provides a custom wrapper that transposes the tensor dimensions (batch, channels, time) so that normalization is applied across channels rather than the time dimension, which is the default PyTorch behavior.

Why does it recommend torch.quantile over kthvalue?

torch.quantile handles floating-point interpolation automatically and accepts float percentiles directly, preventing the common off-by-one and rounding errors associated with the integer-based kthvalue method.

PyTorch Pitfalls & Best Practices

Name: PyTorch Pitfalls & Best Practices
Author: smith6jt-cop

bysmith6jt-cop

0•

데이터 과학 및 ML

Identifies and resolves common PyTorch implementation errors in reinforcement learning and neural network architectures.

This skill provides domain-specific guidance for developers working with PyTorch, specifically focusing on advanced components like Reinforcement Learning (RL) buffers, convolutional neural networks, and robust data validation. It streamlines the debugging process by automatically identifying architectural pitfalls—such as incorrect LayerNorm dimensions for Conv1d or off-by-one errors in percentile calculations—and offering verified implementation patterns to ensure numerical stability and code correctness in production environments.

주요 기능

01Implements dimension-aware LayerNorm wrappers for Conv1d layers

02Automates NaN/Inf validation and minimum sample size checks for calibration

03Corrects percentile calculation logic using torch.quantile instead of kthvalue

04Provides robust bounds-checking for replay buffer sampling and transitions

050 GitHub stars

06Refines action gating logic using explicit torch.where conditional patterns

사용 사례

01Implementing robust data preprocessing and validation pipelines for ML models

02Building stable Reinforcement Learning environments and replay buffers

03Designing complex CNN architectures with specialized normalization layers

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add smith6jt-cop/skills_registry pytorch-common-pitfalls

For use in Claude.ai and ChatGPT

Download Skill