This repository contains all source code, pre-trained models related to the paper "An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator"

Last update: Nov 15, 2021

Overview

An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator

This is a Pytorch implementation for the paper "An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator".

Requirement

python 3.7.3
pytorch 1.2.0
tensorflow 2.0.0
torchtext 0.4.0
torchvision 0.4.0
mnist

Data preparation

Run 0_prepare_dataset.py to automatically download, process and store images of CIFAR10, STL10, MNIST datasets. CAT dataset can be downloaded here.
Download FID stats for CIFAR-10 at http://bioinf.jku.at/research/ttur/ttur_stats/fid_stats_cifar10_train.npz

Training

Run 1_train.sh to train our proposed loss function RMCosGAN along with other loss functions on four datasets.

Appendix

Network Architectures

DCGAN Architecture for CIFAR-10, MNIST and STL-10 datasets

Operation	Filter Units	Non Linearity	Normalization
Generator G(z)
Linear	512	None	None
Trans.Conv2D	256	ReLU	Batch
Trans.Conv2D	128	ReLU	Batch
Trans.Conv2D	64	ReLU	Batch
Trans.Conv2D	3	Tanh	None
Discriminator D(x)
Conv2D	64	Leaky-ReLU	Spectral
Conv2D	64	Leaky-ReLU	Spectral
Conv2D	128	Leaky-ReLU	Spectral
Conv2D	128	Leaky-ReLU	Spectral
Conv2D	256	Leaky-ReLU	Spectral
Conv2D	256	Leaky-ReLU	Spectral
Conv2D	512	Leaky-ReLU	Spectral

DCGAN Architecture for CAT dataset

Operation	Filter Units	Non Linearity	Normalization
Generator G(z)
Trans.Conv2D	1024	ReLU	Batch
Trans.Conv2D	512	ReLU	Batch
Trans.Conv2D	256	ReLU	Batch
Trans.Conv2D	128	ReLU	Batch
Trans.Conv2D	3	Tanh	None
Discriminator D(x)
Conv2D	128	Leaky-ReLU	Spectral
Conv2D	256	Leaky-ReLU	Spectral
Conv2D	512	Leaky-ReLU	Spectral
Conv2D	1024	Leaky-ReLU	Spectral

Experimental results

60 randomly-generated images with RMCosGAN at FID=31.34 trained on CIFAR-10 dataset

60 randomly-generated images with RMCosGAN at FID=13.17 trained on MNIST dataset

60 randomly-generated images with RMCosGAN FID=52.16 trained on STL-10 dataset

60 randomly-generated images with RMCosGAN at FID=9.48 trained on CAT dataset

Citation

Please cite our paper if RMCosGAN is used:

@article{RMCosGAN,
  title={An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator},
  author={Cuong Nguyen, Tien-Dung Cao, Tram Truong-Huu, Binh T.Nguyen},
  journal={},
  year={}
}

If this implementation is useful, please cite or acknowledge this repository on your work.

Contact

Cuong Nguyen ([email protected]),

Tien-Dung Cao ([email protected]),

Tram Truong-Huu ([email protected]),

Binh T.Nguyen ([email protected])

This repository contains all source code, pre-trained models related to the paper "An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator"

Related tags

Overview

An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator

Requirement

Data preparation

Training

Appendix

Network Architectures

Experimental results

Citation

Contact

Owner

Cuong Nguyen

Yolov5 deepsort inference，使用YOLOv5+Deepsort实现车辆行人追踪和计数，代码封装成一个Detector类，更容易嵌入到自己的项目中

Python package to add text to images, textures and different backgrounds

This tutorial aims to learn the basics of deep learning by hands, and master the basics through combination of lectures and exercises

Code for "Solving Graph-based Public Good Games with Tree Search and Imitation Learning"

[ACMMM 2021 Oral] Enhanced Invertible Encoding for Learned Image Compression

PyTorch implementation of "Conformer: Convolution-augmented Transformer for Speech Recognition" (INTERSPEECH 2020)

Emblaze - Interactive Embedding Comparison

This code is the implementation of the paper "Coherence-Based Distributed Document Representation Learning for Scientific Documents".

Code for "PV-RAFT: Point-Voxel Correlation Fields for Scene Flow Estimation of Point Clouds", CVPR 2021

A GPT, made only of MLPs, in Jax

DABO: Data Augmentation with Bilevel Optimization

Educational 2D SLAM implementation based on ICP and Pose Graph

Image reconstruction done with untrained neural networks.

Official implementation of "GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generators" (NeurIPS 2020)

Official implementation of "Learning Not to Reconstruct" (BMVC 2021)

The lightweight PyTorch wrapper for high-performance AI research. Scale your models, not the boilerplate.

Attention-guided gan for synthesizing IR images

FrankMocap: A Strong and Easy-to-use Single View 3D Hand+Body Pose Estimator

(to be released) [NeurIPS'21] Transformers Generalize DeepSets and Can be Extended to Graphs and Hypergraphs

We present a framework for training multi-modal deep learning models on unlabelled video data by forcing the network to learn invariances to transformations applied to both the audio and video streams.