Which data formats does this skill support?

It supports various 'domain formats' including Nixtla (for forecasting), sparse matrices (for tree models), tensors (for PyTorch), and HuggingFace Datasets.

What is the primary role of the Algorithm layer in Haipipe?

The Algorithm layer (Layer 1) represents the raw computational engine, either as an external library or a custom nn.Module, focused strictly on model structure and forward passes.

When should I write custom Layer 1 code versus using a Tuner?

You write Layer 1 code when creating custom architectures like fusion layers. For pure external libraries like XGBoost, you skip Layer 1 code and wrap the library directly in a Layer 2 Tuner.

Can I include training loops in an algorithm_*.py file?

No. Layer 1 must be stateless regarding training; all training logic, optimizers, and loss computations belong in the Tuner layer (Layer 2).

Haipipe NN Algorithm Layer

Name: Haipipe NN Algorithm Layer
Author: jluo41

byjluo41

0•

Ciencia de Datos y ML

Standardizes the integration of external machine learning libraries and custom neural network modules within the Haipipe architecture.

This skill provides a structured framework for Layer 1 of the 4-layer Haipipe neural network pipeline, specifically focusing on the 'Algorithm' layer. It defines patterns for wrapping third-party ML libraries such as XGBoost, PyTorch, and Nixtla, as well as guidelines for creating custom nn.Module subclasses for specialized architectures. By enforcing a strict separation between raw computational logic and pipeline orchestration, this skill enables researchers to swap algorithms seamlessly and maintain clean, modular codebases for complex machine learning projects.

Características Principales

01Framework for developing custom PyTorch nn.Module subclasses (algorithm_*.py)

02Unified serialization guidance for JSON, Pickle, state_dict, and save_pretrained formats

030 GitHub stars

04Standardized patterns for wrapping external libraries like XGBoost, Scikit-learn, and HuggingFace

05Support for diverse domain formats including Nixtla DataFrames, sparse matrices, and tensors

06Architectural enforcement of stateless model definitions decoupled from training logic

Casos de Uso

01Integrating new time-series forecasting libraries into an existing training pipeline

02Standardizing data input and model saving protocols across a team of data scientists

03Adding custom embedding or fusion layers to HuggingFace models for specialized research

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add jluo41/research-skills haipipe-nn-1-algorithm

For use in Claude.ai and ChatGPT

Download Skill