nnirp/NNIRP-dataset
收藏资源简介:
NNIRP数据集是一个用于从ONNX计算图中预测神经网络模型推理运行时的数据集和评估协议。它包含103,070个分析样本,来自6个架构家族的190个源配置,组织成28个子家族的156个集群。每个样本包括三个数据层:分析数据(.json,包含运行时、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,仅包含拓扑结构的轻量级ONNX计算图)。数据集结构按源配置和数据层组织,包含一个tar.gz存档。数据分割采用集群原子分割,确保没有集群跨越两个分割。数据集还提供了详细的PyG特征和分析JSON的模式描述,以及如何加载和使用数据集的示例。数据收集通过一个三阶段的自动化流程完成,包括ONNX导出、GPU分析和特征编码。数据集的主要局限性包括所有分析均在NVIDIA T4 GPU上进行,ONNX导出覆盖不完整,以及运行时测量反映PyTorch eager-mode推理。数据集采用CC-BY-NC-SA 4.0许可证。
The NNIRP dataset is a dataset and evaluation protocol for predicting inference runtime of neural network models from their ONNX computational graphs. It contains 103,070 profiling samples from 190 source configurations spanning 6 architecture families, organized into 156 clusters across 28 sub-families. Each sample includes three data layers: profiling data (.json, containing runtime, VRAM, and RAM statistics), 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). The dataset is organized as one tar.gz archive per source configuration per data layer. Data splits are cluster-atomic, ensuring no cluster straddles two splits. The dataset also provides detailed schemas for PyG features and profiling JSONs, along with examples of how to load and use the dataset. Data was collected through a three-stage automated pipeline: ONNX export, GPU profiling, and feature encoding. Key limitations include profiling performed on a single GPU type (NVIDIA T4), incomplete ONNX export coverage, and runtime measurements reflecting PyTorch eager-mode inference. The dataset is licensed under CC-BY-NC-SA 4.0.




