xuzhuo0417/openacm-gnn-8bit-dataset

8-bit GNN Multiplier Configuration PPA Dataset

This repository contains the 8-bit approximate multiplier configuration dataset used by the OpenACM PEA-GNN surrogate, together with the final model weights and helper code needed for reproduction.

Hugging Face Dataset File

• —data/configs_ppa.csv: the main table for datasets.load_dataset. Each row contains sample_id, the multiplier configuration vector, normalized targets, and raw targets (mred, nmed, delay, area, power).

Example:

from datasets import load_dataset

dataset = load_dataset("xuzhuo0417/openacm-gnn-8bit-dataset")
print(dataset["train"][0])

Repository Layout

• —raw/dir_vectors.txt: original 50000 x 9 configuration vectors.
• —raw/label.txt: normalized 50000 x 5 labels used for GNN training.
• —raw/merged.log: unnormalized 50000 x 5 target values.
• —raw/label_minmax.txt: two-row min/max file used to normalize raw/merged.log into raw/label.txt.
• —raw/Graph.csv: fixed graph adjacency matrix for the multiplier graph.
• —metadata/FEATURE.csv: static feature table retained for reproducibility.
• —metadata/graph_edges.csv: edge list exported from raw/Graph.csv; duplicate edges are preserved when the adjacency entry is greater than 1.
• —metadata/dataset_summary.json: compact machine-readable metadata.
• —code/data_pyg.py and code/node_feature.py: scripts for rebuilding the PyTorch Geometric dataset from raw/.
• —code/merge.py and code/minmax.py: scripts documenting target merge and min-max normalization.
• —model/best_model_weights_8.pth: final GNN model weights.
• —model/gnn_predictor.py and model/my_io.py: inference helper files.
• —model/FEATURE.csv and model/label_minmax_8.txt: model-side metadata used by the inference helpers.

Row Ranges

The main HF table keeps all rows in one file. The row ranges used by the local training scripts are:

all: 0:50000 (50000 rows)
train: 0:46000 (46000 rows)
validation: 46000:48000 (2000 rows)
test: 48000:50000 (2000 rows)

Normalization

raw/label.txt is generated from raw/merged.log using min-max normalization:

target_normalized = (target_raw - target_min) / (target_max - target_min + 1e-12)

The min and max values are stored in raw/label_minmax.txt.

Model Weights

The model/ directory contains final model parameters only, without optimizer or epoch state. For model discovery and download statistics on Hugging Face, these files can also be uploaded to a separate model repository and linked to the same paper page.