File size: 10,060 Bytes
0281c63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
---
license: mit
datasets:
- colabfit/MD22_buckyball_catcher
- colabfit/MD22_AT_AT
- colabfit/MD22_stachyose
- colabfit/MD22_AT_AT_CG_CG
- colabfit/MD22_Ac_Ala3_NHMe
- colabfit/MD22_DHA
- colabfit/MD22_double_walled_nanotube
- yairschiff/qm9
- maomlab/Molecule3D
metrics:
- mae
tags:
- equivariant
- graph neural network
- molecular property prediction
---

# GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks

<div align="center">

[![Paper](https://img.shields.io/badge/Paper-ICLR%202025-blue)](https://openreview.net/pdf?id=5wxCQDtbMo)
[![Project Page](https://img.shields.io/badge/Project-Website-green)](https://www.sarpaykent.com/publications/gotennet/)
[![License](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
[![PyPI - Version](https://img.shields.io/pypi/v/gotennet)](https://pypi.org/project/gotennet/)
[![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-red.svg)](https://pytorch.org/)

</div>

<p align="center">
  <img src="https://raw.githubusercontent.com/sarpaykent/GotenNet/refs/heads/main/assets/GotenNet_framework.png" width="800">
</p>

## Overview

This is the official implementation of **"GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks"** published at ICLR 2025.

GotenNet introduces a novel framework for modeling 3D molecular structures that achieves state-of-the-art performance while maintaining computational efficiency. Our approach balances expressiveness and efficiency through innovative tensor-based representations and attention mechanisms.

## Table of Contents
  - [✨ Key Features](#-key-features)
  - [πŸš€ Installation](#-installation)
    - [πŸ“¦ From PyPI (Recommended)](#-from-pypi-recommended)
    - [πŸ”§ From Source](#πŸ”§-from-source)
  - [πŸ”¬ Usage](#πŸ”¬-usage)
    - [Using the Model](#using-the-model)
    - [Loading Pre-trained Models Programmatically](#loading-pre-trained-models-programmatically)
    - [Training a Model](#training-a-model)
    - [Testing a Model](#testing-a-model)
    - [Configuration](#configuration)
  - [🀝 Contributing](#-contributing)
  - [πŸ“š Citation](#-citation)
  - [πŸ“„ License](#-license)
  - [Acknowledgements](#acknowledgements)

## ✨ Key Features

- πŸ”„ **Effective Geometric Tensor Representations**: Leverages geometric tensors without relying on irreducible representations or Clebsch-Gordan transforms
- 🧩 **Unified Structural Embedding**: Introduces geometry-aware tensor attention for improved molecular representation
- πŸ“Š **Hierarchical Tensor Refinement**: Implements a flexible and efficient representation scheme
- πŸ† **State-of-the-Art Performance**: Achieves superior results on QM9, rMD17, MD22, and Molecule3D datasets
- πŸ“ˆ **Load Pre-trained Models**: Easily load and use pre-trained model checkpoints by name, URL, or local path, with automatic download capabilities.

## πŸš€ Installation

### πŸ“¦ From PyPI (Recommended)

You can install it using pip:

*   **Core Model Only:** Installs only the essential dependencies required to use the `GotenNet` model.
    ```bash
    pip install gotennet
    ```

*   **Full Installation (Core + Training/Utilities):** Installs core dependencies plus libraries needed for training, data handling, logging, etc.
    ```bash
    pip install gotennet[full]
    ```

### πŸ”§ From Source

1.  **Clone the repository:**
    ```bash
    git clone https://github.com/sarpaykent/gotennet.git
    cd gotennet
    ```

2.  **Create and activate a virtual environment** (using conda or venv/uv):
    ```bash
    # Using conda
    conda create -n gotennet python=3.10
    conda activate gotennet

    # Or using venv/uv
    uv venv --python 3.10
    source .venv/bin/activate
    ```

3.  **Install the package:**
    Choose the installation type based on your needs:

    *   **Core Model Only:** Installs only the essential dependencies required to use the `GotenNet` model.
        ```bash
        pip install .
        ```

    *   **Full Installation (Core + Training/Utilities):** Installs core dependencies plus libraries needed for training, data handling, logging, etc.
        ```bash
        pip install .[full]
        # Or for editable install:
        # pip install -e .[full]
        ```
    *(Note: `uv` can be used as a faster alternative to `pip` for installation, e.g., `uv pip install .[full]`)*

## πŸ”¬ Usage

### Using the Model

Once installed, you can import and use the `GotenNet` model directly in your Python code:

```python
from gotennet import GotenNet

# --- Using the base GotenNet model ---
# Requires manual calculation of edge_index, edge_diff, edge_vec

# Example instantiation 
model = GotenNet(
    n_atom_basis=256,
    n_interactions=4,
    # resf of the parameters
)

# Encoded representations can be computed with
h, X = model(atomic_numbers, edge_index, edge_diff, edge_vec) 

# --- Using GotenNetWrapper (handles distance calculation) ---
# Expects a PyTorch Geometric Data object or similar dict
# with keys like 'z' (atomic_numbers), 'pos' (positions), 'batch'

# Example instantiation
from gotennet import GotenNetWrapper
wrapped_model = GotenNetWrapper(
    n_atom_basis=256,
    n_interactions=4,
    # rest of the parameters
)

# Encoded representations can be computed with
h, X = wrapped_model(data) 

```

### Loading Pre-trained Models Programmatically

You can easily load pre-trained `GotenModel` instances programmatically using the `from_pretrained` class method. This method can accept a model alias (which will be resolved to a download URL), a direct HTTPS URL to a checkpoint file, or a local file path. It handles automatic downloading and caching of checkpoints. Pre-trained model weights and aliases are hosted on the [GotenNet Hugging Face Model Hub](https://huggingface.co/sarpaykent/GotenNet).

```python
from gotennet.models import GotenModel

# Example 1: Load by model alias 
# This will automatically download from a known location if not found locally.
# The format is {dataset}_{size}_{target}
model_by_alias = GotenModel.from_pretrained("QM9_small_homo") 

# Example 2: Load from a direct URL
model_url = "https://huggingface.co/sarpaykent/GotenNet/resolve/main/pretrained/qm9/small/gotennet_homo.ckpt" # Replace with an actual URL
model_by_url = GotenModel.from_pretrained(model_url)

# Example 3: Load from a local file path
local_model_path = "/path/to/your/local_model.ckpt" 
model_by_path = GotenModel.from_pretrained(local_model_path)

# After loading, the model is ready for inference:
predictions = model_by_alias(data_input) 
```

For more advanced scenarios, if you only need to load the base `GotenNet` representation module from a local checkpoint (e.g., a checkpoint that only contains representation weights), you can use:

```python
from gotennet.models.representation import GotenNet, GotenNetWrapper

# Example: Load a GotenNet representation from a local file
representation_checkpoint_path = "/path/to/your/local_model.ckpt" 
gotennet_model = GotenNet.load_from_checkpoint(representation_checkpoint_path)
# or
gotennet_wrapped = GotenNetWrapper.load_from_checkpoint(representation_checkpoint_path)
```

### Training a Model

After installation, you can use the `train_gotennet` command:

```bash
train_gotennet
```

Or you can run the training script directly:

```bash
python gotennet/scripts/train.py
```

Both methods use Hydra for configuration. You can reproduce U0 target prediction on the QM9 dataset with the following command:

```bash
train_gotennet experiment=qm9_u0.yaml
```

### Testing a Model

To evaluate a trained model, you can use the `test_gotennet` script. When you provide a checkpoint, the script can infer necessary configurations (like dataset and task details) directly from the checkpoint file. This script leverages the `GotenModel.from_pretrained` capabilities, allowing you to specify the model to test by its alias, a direct URL, or a local file path, handling automatic downloads.

Here's how you can use it:

```bash
# Option 1: Test by model alias (e.g., QM9_small_homo)
# The script will automatically download the checkpoint and infer configurations.
test_gotennet checkpoint=QM9_small_homo

# Option 2: Test with a direct checkpoint URL
# The script will automatically download the checkpoint and infer configurations.
test_gotennet checkpoint=https://huggingface.co/sarpaykent/GotenNet/resolve/main/pretrained/qm9/small/gotennet_homo.ckpt

# Option 3: Test with a local checkpoint file path
test_gotennet checkpoint=/path/to/your/local_model.ckpt
```

The script uses [Hydra](https://hydra.cc/) for any additional or overriding configurations if needed, but for straightforward evaluation of a checkpoint, only the `checkpoint` argument is typically required.

### Configuration

The project uses [Hydra](https://hydra.cc/) for configuration management. Configuration files are located in the `configs/` directory.

Main configuration categories:
- `datamodule`: Dataset configurations (md17, qm9, etc.)
- `model`: Model configurations
- `trainer`: Training parameters
- `callbacks`: Callback configurations
- `logger`: Logging configurations

## 🀝 Contributing

We welcome contributions to GotenNet! Please feel free to submit a Pull Request.


## πŸ“š Citation

Please consider citing our work below if this project is helpful:


```bibtex
@inproceedings{aykent2025gotennet,
  author = {Aykent, Sarp and Xia, Tian},
  booktitle = {The Thirteenth International Conference on LearningRepresentations},
  year = {2025},
  title = {{GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks}},
  url = {https://openreview.net/forum?id=5wxCQDtbMo},
  howpublished = {https://openreview.net/forum?id=5wxCQDtbMo},
}
```

## πŸ“„ License

This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.

## Acknowledgements

GotenNet is proudly built on the innovative foundations provided by the projects below.
- [e3nn](https://github.com/e3nn/e3nn)
- [PyG](https://github.com/pyg-team/pytorch_geometric)
- [PyTorch Lightning](https://github.com/Lightning-AI/pytorch-lightning)