nnirp/NNIRP-dataset-sample

NNIRP数据集是一个用于从ONNX计算图预测神经网络模型推理运行时间的数据集和评估协议。包含约107k个分析样本，来自6个架构家族的190个源配置，分为28个子家族的125个集群。每个样本包括三个数据层：分析数据（.json，包含在NVIDIA T4 GPU上测量的运行时、VRAM和RAM统计信息）、PyG特征（.pt.zst，包含节点、边和图级特征的PyTorch Geometric图编码）和ONNX图（.onnx，轻量级ONNX计算图，仅包含拓扑结构，无训练权重）。数据组织为每个源配置的tar.gz存档，并提供了详细的数据结构、数据分割、特征模式以及加载示例。数据集收集过程包括ONNX导出、GPU分析和特征编码三个阶段。

The NNIRP dataset is a dataset and evaluation protocol for predicting inference runtime of neural network models from their ONNX computational graphs. It contains ~107k profiling samples from 190 source configurations spanning 6 architecture families, organized into 125 clusters across 28 sub-families. Each sample includes three data layers: Profiling (.json, containing runtime, VRAM, and RAM statistics measured on an NVIDIA T4 GPU), PyG Features (.pt.zst, containing PyTorch Geometric graph encodings with node, edge, and graph-level features), and ONNX Graphs (.onnx, lightweight ONNX computational graphs with topology only, no trained weights). The data is organized as one tar.gz archive per source configuration per data layer, with detailed dataset structure, data splits, feature schema, and loading examples provided. The dataset collection process involves a three-stage automated pipeline: ONNX export, GPU profiling, and feature encoding.