# Review — LSGAN: Least Squares Generative Adversarial Networks (GAN)

## Training GAN Using Least Square Loss, Higher Image Quality, Improved Training Stability

In this story, **Least Squares Generative Adversarial Networks**, (LSGAN), by City University of Hong Kong, The Education University of Hong Kong, Northwestern Polytechnical University, and CodeHatch Corp., is briefly reviewed.

**Regular****GAN****s**hypothesize the discriminator as a classifier with the**sigmoid cross entropy loss function**may lead to the**vanishing gradients problem during the learning process.**

In this paper:

**LSGAN**is proposed, in which the**least squares loss function**is adopted for the discriminator.- First, LSGANs are able to
**generate higher quality images**than regular GANs. Second, LSGANs perform**more stable during the learning process**.

This is a paper in **2017 ICCV **with over **2400 citations**. (

# Outline

**Least Squares Generative Adversarial Networks (LSGAN)****Parameters Selection****Experimental Results**

**1. Least Squares Generative Adversarial Networks (LSGAN)**

## 1.1. **Regular **GAN**s**

**Regular****GAN****s**adopt the**sigmoid cross entropy loss function**. When updating the generator, this loss function will cause**the problem of vanishing gradients for the samples that are on the correct side of the decision boundary, but are still far from the real data.**- e.g.: It gets very small errors for the fake samples (in magenta) for updating
*G*as they are on the correct side of the decision boundary.

## 1.2. LSGAN

- To remedy this problem, LSGAN is proposed.
- Suppose we use the
*a*-*b*coding scheme for the discriminator, where*a*and*b*are the labels for fake data and real data**objective functions for LSGANs**can be defined as follows:

- where
*c*denotes the value that G wants D to believe for fake data. - First, the penalization will make the generator to generate samples toward the decision boundary, as shown above.

Moving the generated samples toward the decision boundary leads to making them be closer to the manifold of real data.

Second, penalizing the samples lying a long way to the decision boundary can generate more gradients when updating the generator, which in turn relieves the problem of vanishing gradients.

- This allows LSGANs to perform
**more stable during the learning process.** - As shown above, the least squares loss function is at only at one point, while the sigmoid cross entropy loss function will saturate when
*x*is relatively large.

**2. Parameters Selection**

**One condition**is to set*b-c=*1 and*b*-*a*=2**minimizing the Pearson**between*X*² divergence*pd*+*pg*and 2*pg*.- (Please feel free to read the paper about the relation of LSGAN loss function to f-divergence, particularly the Pearson
*X*² divergence.) - For example, by setting
, we got:*a*=1,*b*=1, and*c*=0

**Another method**is to make*G*generate samples as real as possible by setting. For example, by using the*c*=*b***0-1 binary coding scheme**:

- In the paper, the latter one is used.

**3. Experimental Results**

## 3.1. LSUN Dataset

- The images generated by LSGANs are of
**better quality**than the ones generated by the two baseline methods, DCGANs and EBGANs.

- The results of LSGANs trained on other scene datasets are also shown.

## 3.2. Stability Comparison on LSUN Dataset

- Based on the network architectures of DCGANs, two architectures are designed to compare the stability.
- The first one is to
**exclude the batch normalization for the generator (**, and the second one is to*BN_G*)**exclude the batch normalization for both the generator and the discriminator (**With two optimizers*BN_GD*).**Adam**and**RMSProp**tested. Thus, there are 4 settings in total. - These models are trained on LSUN bedroom dataset using regular GANs and LSGANs separately.
- First, for BNG with Adam, there is a chance for
**LSGANs**to generate relatively good quality images.**With 10 times of testing, and 5 of them succeeds to generate relatively good quality images.** - But for
**regular****GAN****s**, there is**no successful learning**. Regular GANs suffer from a**severe degree of mode collapse.**

- The generated images by LSGANs and regular GANs are shown above.
**LSGANs generate higher quality images**than regular GANs which have a slight degree of mode collapse. - LSGANs and regular GANs have similar performances for
*BN_G*with RMSProp and*BN_GD*with Adam. Specifically, for,*BN_G*with RMSProp**both LSGANs and regular GANs learn the data distribution successfully.** - For
*BN_GD*with Adam**both**the ones have**a slight degree of mode collapse.** - Last,
**RMSProp performs more stable**than Adam since regular.

## 3.2. Stability Comparison on Gaussian Mixture Distribution Dataset

- On this dataset, both the generator and the discriminator use a simple architecture, containing three fully-connected layers.
**Regular****GAN****s suffer from mode collapse starting at step 15k.**They generate samples around a single valid mode of the data distribution.- But LSGANs learn the Gaussian mixture distribution successfully.

## 3.3. Handwritten Chinese Characters

- A conditional LSGAN model is trained.

- LSGANs learn to generate readable Chinese characters successfully, and some randomly selected characters are shown above.
- The
**generated characters**by LSGANs are**readable**. - And the
**labels**of the generated images through label vectors can be**obtained correctly**, which can be used for further applications such as data augmentation.

## Reference

[2017 ICCV] [LSGAN]

Least Squares Generative Adversarial Networks

## Generative Adversarial Network (GAN)

**Image Synthesis** [GAN] [CGAN] [LAPGAN] [AAE] [DCGAN] [CoGAN] [SimGAN] [BiGAN] [ALI] [LSGAN]**Image-to-image Translation **[Pix2Pix] [UNIT] [CycleGAN] [MUNIT]**Super Resolution** [SRGAN & SRResNet] [EnhanceNet] [ESRGAN]**Blur Detection** [DMENet]**Camera Tampering Detection **[Mantini’s VISAPP’19]**Video Coding** [VC-LAPGAN] [Zhu TMM’20] [Zhong ELECGJ’21]