---
license: apache-2.0
language:
- en
tags:
- computational-graphs
- tensor-compiler
- deep-learning
size_categories:
- 100M<n<1B
pretty_name: Dataset
---

# GraphNet: A Large-Scale Computational Graph Dataset for Tensor Compiler Research

GraphNet is a large-scale dataset of deep learning computation graphs, built as a standard benchmark for tensor compiler optimization.

## Dataset Description

GraphNet contains **4,066** computational graph samples from various models, organized into 4 distinct configurations based on graph characteristics:

- **full_graph** (10 samples): Complete model computation graphs
- **typical_graph** (33 samples): Representative subgraphs capturing common patterns
- **fusible_graph** (1,935 samples): Fusible operator subgraphs
- **sole_op_graph** (2,088 samples): Individual operator graphs

### Data Sources

The computational graphs are extracted from popular frameworks and model repositories, such as timm, transformers, mmseg, mmpose, cosyvoice, nemo, ultralytics, etc.

## Dataset Structure

Each configuration contains a Parquet file with the following schema:

| Field | Type | Description |
|-------|------|-------------|
| `uuid` | string | Unique identifier for the graph sample |
| `repo_name` | string | Source repository name |
| `relative_model_path` | string | Path to model within repository |
| `sample_type` | string | Graph type (full_graph, typical_graph, etc.) |
| `is_subgraph` | boolean | Whether this is extracted from a larger graph |
| `num_ops` | integer | Number of operations in the graph |
| `graph_hash` | string | Unique hash of the graph structure |
| `framework` | string | Framework used (torch) |
| `dynamic` | boolean | Whether graph has dynamic shapes |
| `source` | string | Original model source |
| `heuristic_tag` | string | Domain tag (audio, vision, nlp, etc.) |
| `dimension_generalization_passes` | JSON | Applied dimension generalization passes |
| `data_type_generalization_passes` | JSON | Applied datatype generalization passes |

## Usage

### Load with `datasets` library

```python
from datasets import load_dataset

# Load specific configurations using data_dir
full_graph = load_dataset("PaddlePaddle/GraphNet", data_dir="full_graph")
typical_graph = load_dataset("PaddlePaddle/GraphNet", data_dir="typical_graph")
fusible_graph = load_dataset("PaddlePaddle/GraphNet", data_dir="fusible_graph")
sole_op_graph = load_dataset("PaddlePaddle/GraphNet", data_dir="sole_op_graph")

# Access data
print(f"Full graphs: {len(full_graph['train'])} samples")
print(f"First sample: {full_graph['train'][0]}")
```

### Load all data at once

```python
from datasets import load_dataset

# Load all samples together
dataset = load_dataset("PaddlePaddle/GraphNet")
print(f"Total samples: {len(dataset['train'])}")
```


## Dataset Statistics

- **Total Samples**: 4,066
- **Sample Types**: 4 configurations
- **Domains**: Audio, Vision, NLP, Multi-modal

## Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{graphnet2026,
  title={GraphNet: A Large-Scale Computational Graph Dataset for Tensor Compiler Research},
  author={PaddlePaddle},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/PaddlePaddle/GraphNet}}
}
```

## License

Apache License 2.0 - See LICENSE file for details