DSW

KO | EN

gitlite — search

#python #typescript #llm #machine-learning #rust #deep-learning #linux #docker #windows #awesome #kubernetes #awesome-list #openai #open-source

(★ 51)

Distributional Sliced-Wasserstein distance code

DSW 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

## Distributional Sliced Wasserstein distance
This is a pytorch implementation of distributional sliced Wasserstein which is a sliced optimal transport distance between two probability measures. Details of the model architecture and experimental results can be found in [our following paper](https://arxiv.org/pdf/2002.07367.pdf).
````
@inproceedings{
nguyen2021distributional,
title={Distributional Sliced-Wasserstein and Applications to Generative Modeling},
author={Khai Nguyen and Nhat Ho and Tung Pham and Hung Bui},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=QYjO70ACDK}
}
````
Please CITE our paper whenever this repository is used to help produce published results or incorporated into other software.
### Installation

* python>=3.6
* pytorch>=1.4
* torchvision
* numpy
* tqdm
* POT
* scikit-image

Options for Max Sliced-Wasserstein distance and Distributional distances (number of gradient steps for find the max slice or the optimal push-forward function):

```
--niter=10
```
Options for Distributional Sliced-Wasserstein Distance and Distributional Generalized Sliced-Wasserstein Distance (regularization strength)

````
--lam=10
````
Options for Generalized Wasserstein Distance (using circular function for Generalized Radon Transform)

````
--r=1000;\
--g='circular'
````
### Train on CELEBA and CIFAR10 and LSUN
``` 
python main.py \
    --datadir='./' \
    --outdir='./result' \
    --batch-size=512 \
    --seed=16 \
    --p=2 \
    --lr=0.0005 \
    --model-type='DSWD'\
    --dataset='CELEBA'
    --latent-size=100 \ 
```
```
model-type in (SWD|MSWD|DSWD|GSWD|DGSWD)
```
Options for Sliced distances (number of projections used to approximate the distances)
````
--num-projection=1000
````
Options for Max Sliced-Wasserstein distance and Distributional distances (number of gradient steps for find the max slice or the optimal push-forward function):

```
--niter=1
```
Options for Distributional Sliced-Wasserstein Distance and Distributional Generalized Sliced-Wasserstein Distance (regularization strength)

````
--lam=1
````
Options for Generalized Wasserstein Distance (using circular function for Generalized Radon Transform)

````
--r=1000;\
--g='circular'
````
### Evaluation
Please use https://github.com/bioinf-jku/TTUR for evaluating the trained generative models.

### Some generated images
##### MNIST generated images
![MNIST](https://raw.githubusercontent.com/VinAIResearch/DSW/master/images/MNIST.png) 
##### CELEBA generated images
![MNIST](https://raw.githubusercontent.com/VinAIResearch/DSW/master/images/CelebA.png) 
##### LSUN generated images
![MNIST](https://raw.githubusercontent.com/VinAIResearch/DSW/master/images/LSUN.png)

### Acknowledgement
Our code uses the implementation of Max-SW, GSW, Max-GSW-NN from https://github.com/kimiandj/gsw.

DSW

DSW 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

DSW

DSW 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README