HALO

KO | EN

gitlite — search

#python #hacktoberfest #ai #llm #deep-learning #docker #nodejs #awesome #cli #kubernetes #vue #awesome-list #mcp #openai

(★ 25)

Official PyTorch implementation of the ICML 2024 paper "Hyperbolic Active Learning for Semantic Segmentation under Domain Shift"

HALO Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# HALO

This is the official PyTorch implementation of the **ICML 2024** paper [**Hyperbolic Active Learning for Semantic Segmentation under Domain Shift**](https://arxiv.org/abs/2306.11180).

*Luca Franco &dagger; 1, Paolo Mandica &dagger; 2, Konstantinos Kallidromitis 3, Devin Guillory4, Yu-Teng Li4, Trevor Darrell4, Fabio Galasso1,2* 
*1 ItalAI,2 Sapienza University of Rome, 3 Panasonic North America, 4 UC Berkeley* 
&dagger; Equal contribution

[[`arXiv`](https://arxiv.org/abs/2306.11180)][[`BibTeX`](#Citation)][[`OpenReview`](https://openreview.net/forum?id=hKdJPMQvew)]

---

[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/hyperbolic-active-learning-for-semantic/domain-adaptation-on-gta5-to-cityscapes)](https://paperswithcode.com/sota/domain-adaptation-on-gta5-to-cityscapes?p=hyperbolic-active-learning-for-semantic)
[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/hyperbolic-active-learning-for-semantic/domain-adaptation-on-synthia-to-cityscapes)](https://paperswithcode.com/sota/domain-adaptation-on-synthia-to-cityscapes?p=hyperbolic-active-learning-for-semantic)
[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/hyperbolic-active-learning-for-semantic/source-free-domain-adaptation-on-gta5-to)](https://paperswithcode.com/sota/source-free-domain-adaptation-on-gta5-to?p=hyperbolic-active-learning-for-semantic)
[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/hyperbolic-active-learning-for-semantic/domain-adaptation-on-cityscapes-to-acdc)](https://paperswithcode.com/sota/domain-adaptation-on-cityscapes-to-acdc?p=hyperbolic-active-learning-for-semantic)

## Usage

You can choose to run the code in a conda environment or in a docker container. Follow the instructions below to set up your preferred environment.

### Conda env setup

```bash
conda create --name halo -y python=3.9
conda activate halo
pip install -r requirements.txt
```

### Docker env setup

Modify the `docker-compose.yml` file to set the correct number of available GPUs. Then set your correct path for the `DATA_VOLUME` and run the following commands:

```bash
export UID=$(id -u)
export GID=$(id -g)
export DATA_VOLUME="YOUR_PATH_TO_DATA_FOLDER"

docker compose build
docker compose up -d
```

### Data Preparation

1. Download the [Cityscapes Dataset](https://www.cityscapes-dataset.com/), [GTAV Dataset](https://download.visinf.tu-darmstadt.de/data/from_games/), [SYNTHIA Dataset](https://synthia-dataset.net/), [ACDC Dataset](https://acdc.vision.ee.ethz.ch/) and extract them.

2. If you downloaded the datasets in a different folder, symlink them to the `datasets` directory:

```bash
    ln -s /path_to_cityscapes_dataset datasets/cityscapes
    ln -s /path_to_gtav_dataset datasets/gtav
    ln -s /path_to_synthia_dataset datasets/synthia
    ln -s /path_to_acdc_dataset datasets/acdc
    ```

3. Generate the label static files for GTAV/SYNTHIA Datasets by running

```bash
    python datasets/generate_gtav_label_info.py -d datasets/gtav -o datasets/gtav/
    python datasets/generate_synthia_label_info.py -d datasets/synthia -o datasets/synthia/
    ```

The `datasets` directory tree should be structured as follows:

### Training

The config files for the different training protocols can be found in the `configs` directory.

Before running the training/testing scripts, make sure to set the correct paths in the config files.

```bash
python train.py -cfg CONFIG_PATH
```

### Testing

```bash
python test.py -cfg CONFIG_PATH
```

### Training with ACDC Dataset or SegFormer architecture

To train with the ACDC dataset or the SegFormer architecture, you need to switch to the corresponding branch:

```bash
git checkout acdc
```
or
```bash
git checkout segformer
```

The two branches will be merged into the main branch soon.

## Acknowledgements

This project is based on the [RIPU](https://github.com/BIT-DA/RIPU) open-source project, with our re-implementation in PyTorch Lightning for multi-gpu active learning. We thank the authors for making the source code publically available.

## Licence

This project is licensed under the terms of the MIT license.

## <a name="Citation"></a>Citation

If you find this repository useful, please consider giving a star :star: and citation:

```latex
@inproceedings{
    franco2024hyperbolic,
    title={Hyperbolic Active Learning for Semantic Segmentation under Domain Shift},
    author={Luca Franco and Paolo Mandica and Konstantinos Kallidromitis and Devin Guillory and Yu-Teng Li and Trevor Darrell and Fabio Galasso},
    booktitle={Forty-first International Conference on Machine Learning},
    year={2024},
    url={https://openreview.net/forum?id=hKdJPMQvew}
}
```

HALO

HALO Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

HALO

HALO Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README