metric-flow-matching

KO | EN

gitlite — search

#javascript #python #typescript #llm #go #golang #rust #deep-learning #linux #cli #kubernetes #ios #mcp #openai

metric-flow-matching

(★ 63)

Official implementation of Metric Flow Matching (NeurIPS 2024)

metric-flow-matching Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# [Metric Flow Matching for Smooth Interpolations on the Data Manifold](https://arxiv.org/abs/2405.14780)

[![arxiv](https://img.shields.io/badge/arxiv-blue)](https://arxiv.org/abs/2405.14780)
[![twitter](https://img.shields.io/badge/twitter-thread-green)](https://x.com/KKapusniak1/status/1797632928920564014)

</div>

## Installation

To set up the environment, you need to install the required dependencies. You can do this by using the `requirements.txt` file.

```bash
conda create --name myenv python=3.11
conda activate myenv
pip install -r requirements.txt
```

## Datasets

Please download the following datasets to run the experiments.

- **Lidar**: [Link to Lidar dataset](https://github.com/facebookresearch/generalized-schrodinger-bridge-matching?tab=readme-ov-file)
- **Single Cell**:
  - CITE and Multi: [Link to CITE and Multi datasets](https://data.mendeley.com/datasets/hhny5ff7yj/1)
  - EB: [Link to EB dataset](https://github.com/KrishnaswamyLab/TrajectoryNet/tree/master/data)
- **Animal Faces HQ (AFHQ)**: [Link to AFHQ dataset](https://github.com/clovaai/stargan-v2#animal-faces-hq-dataset-afhq)

## Running Experiments

All hyperparameters used for the experiments in the paper are located in the [`config`](./configs) folder, with specific definitions in [`mfm/train/parsers.py`](./mfm/train/parsers.py). To specify the data location, use the `--working_dir` flag.

To specify the experiment to run use `--config_path` flag, for example:

```bash
python -m mfm.train.main --config_path ./configs/arch/ot-mfm.yaml
```

## Evaluation

For the `arch`, `sphere`, `single cell`, and `images` experiments, evaluation metrics will be logged after training. Plots for `arch`, `lidar`, and `sphere` will also be saved at the end of training in the `--working_dir` folder.

Model checkpoints are saved within the `checkpoints` folder under `--working_dir`. The `geopath` model can be loaded using the `--load_geopath_model_ckpt <checkpoint_path>` flag. Training and evaluation can be resumed from a flow model checkpoint using the `--resume_flow_model_ckpt <checkpoint_path>` flag.

## Citation

If you find this repository helpful for your publications, please consider citing our paper:
```
@article{kapusniak2024metric,
  title={Metric Flow Matching for Smooth Interpolations on the Data Manifold},
  author={Kapusniak, Kacper and Potaptchik, Peter and Reu, Teodora and Zhang, Leo and Tong, Alexander and Bronstein, Michael and Bose, Avishek Joey and Di Giovanni, Francesco},
  journal={arXiv preprint arXiv:2405.14780},
  year={2024}
}
```

## Files Structure
```
mfm
├── dataloaders
│   ├── image_data.py
│   ├── lidar_data.py
│   └── trajectory_data.py
├── flow_matchers
│   ├── ema.py
│   ├── eval_utils.py
│   ├── flow_net_train.py
│   ├── geopath_net_train.py
│   └── models
│       └── mfm.py
├── geo_metrics
│   ├── land.py
│   ├── metric_factory.py
│   └── rbf.py
├── networks
│   ├── flow_networks
│   │   └── mlp.py
│   ├── geopath_networks
│   │   ├── mlp.py
│   │   └── unet.py
│   ├── mlp_base.py
│   ├── unet_base.py
│   └── utils.py
├── train
│   ├── main.py
│   ├── parsers.py
│   └── train_utils.py
└── utils.py
```

metric-flow-matching

metric-flow-matching Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

metric-flow-matching

metric-flow-matching Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README