1 Introduction
Generative adversarial networks (GANs) aim to approximate a data distribution , using a parameterized model distribution . They achieve this by jointly optimizing generative and discriminative networks (Goodfellow et al., 2014)
. GANs are endtoend differentiable. Samples from the generative network are propagated forward to a discriminative network, and error signals are then propagated backwards from the discriminative network to the generative network. The discriminative network is often viewed as a learned, adaptive loss function for the generative network.
GANs have achieved stateoftheart results for a number of applications (Goodfellow, 2016)
, producing more realistic, sharper samples than other popular generative models, such as variational autoencoders
(Kingma & Welling, 2014). Because of their success, many GAN frameworks have been proposed. However, it has been difficult to compare these algorithms and understand their relative strengths and weaknesses because we are currently lacking in quantitative methods for assessing the learned generators.In this work, we propose new metrics for measuring how realistic are samples generated from GANs. These criteria are based on a formulation of divergence between the distributions and (Nowozin et al., 2016; Sriperumbudur et al., 2009):
(1) 
Here, different choices of , , and can correspond to different divergences (Nowozin et al., 2016)
or different integral probability metrics (IPMs)
(Sriperumbudur et al., 2009). Importantly,can be estimated using samples from
and , and does not require us to be able to estimate or for samples . Instead, evaluating involves finding the function that is maximally different with respect to and .This measure of divergence between the distributions and is related to the GAN criterion if we restrict the function class
to be neural network functions parameterized by the vector
and the class of approximating distributions to correspond to neural network generators parameterized by the vector , allowing formulation as a minmax problem:(2) 
In this formulation, corresponds to the generator network’s distribution and corresponds to the discriminator network (see (Nowozin et al., 2016) for details).
We propose using to evaluate the performance of the generator network for various choices of and , corresponding to different divergences or IPMs between distributions and , that have been successfully used for GAN training. Our proposed metrics differ from most existing metrics in that they are adaptive, and involve finding the maximum over discriminative networks. We compare four metrics, those corresponding to the original GAN (GC) (Goodfellow, 2016), the LeastSquares GAN (LS) (Mao et al., 2017),the Wasserstein GAN (IW) (Gulrajani et al., 2017), and the Maximum Mean Discrepency GAN (MMD) (Li et al., 2017) criteria. Choices for , , and for these metrics are shown in Table 1. Our method can easily be extended to other divergences or IPMs.
Metric  Function Class  

GAN (GC)  ,  
LeastSquares GAN (LS)  ,  
MMD  
Wasserstein (IW) 
To compare these and previous metrics for evaluating GANs, we performed many experiments, training and comparing multiple types of GANs with multiple architectures on multiple data sets. We qualitatively and quantitatively compared these metrics to human perception, and found that our proposed metrics better reflected human perception. We also show that rankings produced using our proposed metrics are consistent across metrics, thus are robust to the exact choices of the functions and in Equation 2.
We used the proposed metrics to quantitatively analyze three different families of GANs: Deep Convolutional Generative Adversarial Networks (DCGAN) (Radford et al., 2015), LeastSquares GANs (LSDCGAN), and Wasserstein GANs (WDCGAN), each of which corresponded to a different proposed metric. Interestingly, we found that the different proposed metrics still agreed on the best GAN framework for each dataset. Thus, even though, e.g. for MNIST the WDCGAN was trained with the IW criterion, LSDCGAN still outperformed it based on the IW criterion at test time.
Our analysis also included carrying out a sensitivity analysis with respect to various factors, such as the architecture size, noise dimension, update ratio between discriminator and generator, and number of data points. Our empirical results show that: i) the larger the GAN architecture, the better the results; ii) having a generator network larger than the discriminator network does not yield good results; iii) the best ratio between the discriminator and generator updates depends on the data set; and iv) the WDCGAN and LSDCGAN performance increases much faster than DCGAN as the number of training examples grows. These metrics thus allow us to tune the hyperparameters and architectures of GANs based on our proposed method.
2 Related Work
GANs can be evaluated using manual annotations, but this is time consuming and difficult to reproduce. Several automatically computable metrics have been proposed for evaluating the performance of probabilistic general models and GANs in particular. We review some of these here, and compare our proposed metrics to these in our experiments.
Many previous probabilistic generative models were evaluated based on the pointwise likelihood of the test data, the criterion also used during training. While GANs can be used to generate samples from the approximating distribution, their likelihood on test samples cannot be evaluated without simplifying assumptions. As discussed in (Theis et al., 2015), likelihood often does not provide good rankings of how realistic the samples look, which is the main goal of GANs. We evaluted the efficacy of the loglikelihood of the test data, as estimated using Annealed Importance Sampling (AIS) (Wu et al., 2016). AIS has been to estimate the likelihood of a test sample
by considering many intermediate distributions that are defined by taking a weighted geometric mean between the prior (input) distribution,
, and an approximation of the joint distribution
. Here,is a Gaussian kernel with fixed standard deviation
around mean . The final estimate depends critically on the accuracy of this approximation. In Section 4, we demonstrate that the AIS estimate ofis highly dependent on the choice of this hyperparameter.
The Generative Adversarial Metric (Im et al., 2016a) measures the relative performance of two GANs by measuring the likelihood ratio of the two models. Consider two GANs with their respective trained partners, and , where and are the generators and and are the discriminators. The hypothesis is that is better than if fools more than fools , and vice versa for the hypothesis . The likelihoodratio is defined as:
(3) 
where and are the swapped pairs and , and is the likelihood of generated from the data distribution under model and indicates that discriminator thinks is a real sample. To evaluate this, we measure the ratio of how frequently , the generator from model 1, fools , the discriminator from model 2, and viceversa: , where and . There are two main caveats to the Generative Adversarial Metric. First, the measurement only provides comparisons between pairs of models. Second, the metric has a constraint where the two discriminators must have an approximately similar performance on a calibration dataset, which can be difficult to satisfy in practice.
The Inception Score (Salimans et al., 2016)
(IS) measures the performance of a model using a thirdparty neural network trained on a supervised classification task, e.g. ImageNet. The IS computes the expectation of divergence between the distribution of class predictions for samples from the GAN compared to the distribution of class labels used to train the thirdparty network,
(4) 
Here, the class prediction given a sample is computed using the thirdparty neural network. In (Salimans et al., 2016), Google’s Inception Network (Szegedy et al., 2015)
trained on ImageNet was the thirdparty neural network. IS is the most widely used metric to measure GAN performance. However, summarizing samples as the class prediction from a network trained for a different task discards much of the important information in the sample. In addition, it requires another neural network that is trained separately via supervised learning. We demonstrate an example of a failure case of IS in the Experiments section.
The Fréchet Inception Distance (FID) (Heusel et al., 2017) extends upon IS. Instead of using the final classification outputs from the thirdparty network as representations of samples, it uses a representation computed from a late layer of the thirdparty network. It compares the mean and covariance of the Inceptionbased representation of samples generated by the GAN to the mean and covariance of the same representation for training samples:
(5) 
This method relies on the Inceptionbased representation of the samples capturing all important information and the first two moments of the distributions being descriptive of the distribution.
Classifier TwoSample Tests (C2ST) (LopezPaz & Oquab, 2016)
proposes training a classifier, similar to a discriminator, that can distinguish real samples from
from generated samples from , and using the error rate of this classifier as a measure of GAN performance. In their work, they used singlelayer andnearest neighbor (KNN) classifiers trained on a representation of the samples computed from a late layer of a thirdparty network (in this case, ResNet
(He et al., 2015)). C2ST is an IPM (Sriperumbudur et al., 2009), like the MMD and Wasserstein metrics we propose, with and , but with a different function class , corresponding to the family of classifiers chosen (in this case, singlelayer networks or KNN, see see our detailed explanation in Appendix Relationship between metrics and binary classification). The accuracy of a classifier trained to distinguish samples from distributions and is just one way to measure the distance between these distributions, and, in this work, we propose a general family.3 Evaluation Metrics
Given a generator with parameters which generates samples from the distribution , we propose to measure the quality of by estimating divergence between the true data distribution and for different choices of divergence measure. We train both and on a training data set, and measure performance on a separate test set. See Algorithm 1 for details. We consider metrics from two widely studied divergence and distance measures, divergence (Nguyen et al., 2008) and the Integral Probability Metric (IPM) (Muller, 1997).
In our experiments, we consider the following four metrics that are commonly used to train GANs. Below, represents the parameters of the discriminator network and represents the parameters of the generator network.
Original GAN Criterion (GC)
Training a standard GAN corresponds to minimizing the following (Goodfellow et al., 2014):
(6) 
where is the prior distribution of the generative network and is a differentiable function from to the data space represented by a neural network with parameter .
is trained with a sigmoid activation function, thus its output is guaranteed to be positive.
LeastSquares GAN Criterion (LS)
A LeastSquares GAN corresponds to training with a Pearson divergence (Mao et al., 2017):
(7) 
Following (Mao et al., 2017), we set and when training .
Maximum Mean Discrepancy (MMD) The maximum mean discrepancy metric considers the largest difference in the expectations over a unit ball of RKHS ,
(8)  
(9) 
where is the RKHS with kernel (Gretton et al., 2012). In this case, we do not need to train a discriminator to evaluate our metric.
Improved Wasserstein Distance (IW)
4 Experiments
The goals in our experiments are twofold. First, we wanted to evaluate the metrics we proposed for evaluating GANs. Second, we wanted to use these metrics to evaluate GAN frameworks and architectures. In particular, we evaluated how the size of the discriminator and generator networks affected performance, and the sensitivity of each algorithm to training data set size.
GAN frameworks. We conducted our experiments on three types of GANs: Deep Convolutional Generative Adversarial Networks (DCGAN), LeastSquares GANs (LSDCGAN), and Wasserstein GANs (WDCGAN). Note that to not confuse the test metric names with the GAN frameworks we evaluated, we use different abbreviations. GC is the original GAN criterion, which is used to train DCGANs. The LS criterion is used to train the LSDCGAN, and the IW is used to train the WDCGAN.
Evaluation criteria. We evaluated these three families of GANs with six metrics. We compared our four proposed metrics to the two most commonly used metrics for evaluating GANs, the IS and FID. Because the optimization of a discriminator is required both during training and test time, we will call the discriminator learned for evaluation of our metrics the critic, in order to not confuse the two discriminators.
We also compared these metrics to human perception, and had three volunteers evaluate and compare sets of images, either from the training data set or generated from different GAN frameworks during training.
Data sets. In our experiments, we considered the MNIST (LeCun et al., 1998), CIFAR10, LSUN Bedroom, and Fashion MNIST datasets. MNIST consists of 60,000 training and 10,000 test images with a size of 28 28 pixels, containing handwritten digits from the classes 0 to 9. From the 60,000 training examples, we set aside 10,000 as validation examples to tune various hyperparameters. Similarly, FashionMNIST consists exactly the same number of training and test examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. The CIFAR10 dataset^{1}^{1}1https://github.com/Lasagne/Recipes/blob/master/papers/deep_residual_learning/Deep_Residual_Learning_CIFAR10.py consists of images with a size of 32 32 3 pixels, with ten different classes of objects. We used 45,000, 5,000, and 10,000 examples as training, validation, and test data, respectively. The LSUN Bedroom dataset consists of images with a size of 6464 pixels, depicting various bedrooms. From the 3,033,342 images, we used 90,000 images as training data and 90,000 images as validation data. The learning rate was selected from discrete ranges and chosen based on a heldout validation set.
Hyperparameters. Table 10 in the Appendix shows the learning rates and the convolutional kernel sizes that were used for each experiment. The architecture of each network is presented in the Appendix in Figure 10
. Additionally, we used exponentialmeansquare kernels with several different sigma values for MMD. A pretrained logistic regression and pretrained residual network were used for IS and FID on the MNIST and CIFAR10 datasets, respectively. For every experiment, we retrained 10 times with different random seeds, and report the mean and standard deviation.
4.1 Qualitative Observations about Existing Metrics
The loglikelihood measurement is the most commonly used metric for generative models. We measured the loglikelihood using AIS^{2}^{2}2We used the original source code from https://github.com/tonywu95/eval_gen on GANs, as shown in Figure 2
. We measured the loglikelihood of the DCGAN on MNIST with three different variances,
, , and. The figure illustrates that the loglikelihood curve over the training epochs varies substantially depending on the variance, which indicates that the fixed Gaussian observable model might not be the ideal assumption for GANs. Moreover, we observe a high loglikelihood at the beginning of training, followed by a drop in likelihood, which then returns to the high value.
The IS and MMD metrics do not require training a critic. It was easy to find samples for which IS and MMD scores did not match their visual quality. For example, Figure 2 shows samples generated by a DCGAN when it failed to train properly. Even though the failed DCGAN samples are much darker than the samples on the right, the IS for the left samples is higher/better than for the right samples. As the ImageNettrained network is likely trained to be somewhat invariant to overall intensity, this issue is to be expected.
A failure case for MMD is shown in Figure 5. The samples on the right are dark, like the previous examples, but still textually recognizable, whereas the samples on the left are totally meaningless. However, MMD gives lower/worse distances to the left samples. The average intensity of the pixels of the left samples are closer to that for the training data, suggesting that MMD is overly sensitive to image intensity. Thus, IS is undersensitive to image intensity, while MMD if oversensitive to it. In Section 4.2.1, we conduct more systematic experiments by measuring the correlation between these metrics to human perceptual scores.
4.2 Metric comparison
Model  MMD  IW  GC  LS  IS 

(Logistic Reg.)  
DCGAN  0.028 0.0066  7.01 1.63  2.2e3 3e4  0.12 0.013  5.76 0.10 
WDCGAN  0.006 0.0009  7.71 1.89  4e4 4e4  0.05 0.008  5.17 0.11 
LSDCGAN  0.012 0.0036  4.50 1.94  3e3 6e4  0.13 0.022  6.07 0.08 
Model  MMD  IW  LS  IS  FID 

(ResNet)  
DCGAN  0.0538 0.014  8.844 2.87  0.0408 0.0039  6.649 0.068  0.112 0.010 
WDCGAN  0.0060 0.001  9.875 3.42  0.0421 0.0054  6.524 0.078  0.095 0.003 
LSDCGAN  0.0072 0.0024  7.10 2.05  0.0535 0.0031  6.761 0.069  0.088 0.008 
Model  MMD  IW  LS 

DCGAN  0.00708  3.79097  0.14614 
WDCGAN  0.00584  2.91787  0.20572 
LSDCGAN  0.00973  3.36779  0.17307 
Model  MNIST  FashionMNIST  

IW  LS  FID  IW  LS  FID  
DCGAN  0.4814 0.0083  0.111 0.0074  1.84 0.15  0.69 0.0057  0.0202 0.00242  3.23 0.34 
EBGAN  0.7277 0.0159  0.029 0.0026  5.36 0.32  0.99 0.0001  2.2e5 5.3e5  104.08 0.56 
WDCGAN GP  0.7314 0.0194  0.035 0.0059  2.67 0.15  0.89 0.0086  0.0005 0.00037  2.56 0.25 
LSDCGAN  0.5058 0.0117  0.115 0.0070  2.20 0.27  0.68 0.0086  0.0208 0.00290  0.62 0.13 
BEGAN    0.009 0.0063  15.9 0.48  0.90 0.0159  0.0016 0.00047  1.51 0.16 
DRAGAN  0.4632 0.0247  0.116 0.0116  1.09 0.13  0.66 0.0108  0.0219 0.00232  0.97 0.14 
To both compare the metrics as well as different GAN frameworks, we evaluated the six metrics on different GAN frameworks. Tables 5, 5, and 5 present the results on MNIST, CIFAR10, and LSUN respectively.
As each type of GAN was trained using one of our proposed metrics, we investigated whether the metric favors samples from the model trained using the same metric. Interestingly, we do not see this behavior, and our proposed metrics agree on which GAN framework produces samples closest to the test data set. Every metric, except for MMD, showed that LSDCGAN performed best for MNIST and CIFAR10, while WDCGAN performed best for LSUN. As discussed below, we found DCGAN to be unstable to train, and thus excluded GC as a metric for experiments except for this first data set. For FashionMNIST, FID’s ranking disagreed with IW and LS.
We observed similar results for a range of different critic CNN architectures (number of feature maps in each convolutional layer): , , , and (see Supp. Fig. 13 and 13).
We evaluated a larger variety of GAN frameworks using pretrained GANs downloaded from (pyt, ). In particular, we evaluated on EBGAN (Junbo Zhao, 2016), BEGAN (Berthelot et al., 2017), WDCGAN GP (Gulrajani et al., 2017), and DRAGAN (Kodali et al., 2017). Table 5 presents the evaluation results. Critic architectures were selected to match those of these pretrained GANs. For both MNIST and FashionMNIST, the three metrics are consistent and they rank DRAGAN the highest, followed by LSDCGAN and DCGAN.
The standard deviations for the IW distance are higher than for LS divergence. We computed the Wilcoxon rank sum in order to test that whether medians of the distributions of distances are the same for DCGAN, LSDCGAN, and WDCGAN. We found that the different GAN frameworks have significantly different performance according to the LSGAN criterion, but not according to the IW criterion (, Wilcoxon ranksum test). Thus LS is more sensitive than IW.
We evaluated the consistency of the metrics with respect to the size of the validation set. We trained our three GAN frameworks for 100 epochs with training 90,000 examples from the LSUN Bedroom dataset. We then trained LS and IW critics using both 300 and 90,000 validation examples. We looked at how often the critic trained with 300 examples agreed with that trained with 90,000 examples. The LS critics agreed 88% of the time, while the IW critics agreed only 55% of the time (slightly better than chance). Thus, LS is more robust to validation data set size. Another advantage is that measuring the LS distance is faster than measuring the IW distance, as estimating IW involves regularizing with a gradient penalty (Gulrajani et al., 2017). Computing the gradient penalty term and tuning its regularization coefficient requires extra computational time.
As mentioned above, we found training a critic using the GC criterion (corresponding to a DCGAN) to be unstable. It has previously been speculated that this is the case because the support of the data and model distributions possibly becoming disjoint (Arjovsky & Bottou, 2017), and the Hessian of the GAN objective being nonHermitian (Mescheder et al., 2017). LSDCGAN and WDCGAN propose to address this by providing nonsaturating gradients. We also found DCGAN to be difficult to train, and thus only report results using the corresponding criterion GC for MNIST. Note that this is different than training a discriminator as part of standard GAN training because we are training from a random initialization, not from the previous version of the discriminator.
Our experience was that the LSDCGAN was the simplest and most stable model to train. We visualized the 2D subspace of the loss surface of the GANs in Supp. Fig. 29
. Here, we took the parameters of three trained models (corresponds to red vertices in the figure) and applied barycentric interpolation with respect to three parameters (see details from
(Im et al., 2016c)). DCGAN surfaces have much sharper slopes when compared to the LSDCGAN and WDCGAN, and LSDCGAN has the most gentle surfaces. In what follows, we show that this geometric view is consistent with our finding that LSDCGAN is the easiest and the most stable to train.4.2.1 Comparison to Human Perception
We compared the LS, IW, MMD, and IS metrics to human perception for the CIFAR10 dataset. To accomplish this, we asked five volunteers to choose which of two sets of 100 samples, each generated using a different generator, looked most realistic. Before surveying, the volunteers were trained to choose between real samples from CIFAR10 and samples generated by a GAN. Supp. Fig. 15 displays the user interface for the participants, and Supp. Fig. 15 shows the fraction of labels that the volunteers agreed upon.
Table 6) presents the fraction of pairs for which each metric agrees with humans (higher is better). IW has a slight edge over LS, and both outperform IS and MMD. In Figure 3, we show examples in which all humans agree and metrics disagree with human perception. All such examples are shown in Supp. Fig. 2124.
Metric  Fraction  [Agreed/Total] samples  p < .05? 

IW  0.977  128 / 131  * * 
LS  0.931  122 / 131  * 
IS  0.863  113 / 131  * 
MMD  0.832  109 / 131  * * 
4.3 Sensitivity Analysis
4.3.1 Performance change with respect to the size of the network
Several works have demonstrated an improvement in performance by enlarging deep network architectures (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; He et al., 2015; Huang et al., 2017). Here, we investigate performance changes with respect to the width and depth of the networks.
First, we trained three GANs with varying numbers of feature map sizes, as shown in Table 7 (ad). Note that we double the number of feature maps in Table 7 for both the discriminators and generators. In Figure 5, the performance of the LS score increases logarithmically as the number of feature maps is doubled. A similar behaviour is observed in other metrics as well (see S.M. Figure 17).
Label  Feature Maps  
Discriminator  Generator  
(a)  [3, 16 , 32 , 64 ]  [128 , 64 , 32 , 3] 
(b)  [3, 32 , 64 , 128]  [256 , 128, 64 , 3] 
(c)  [3, 64 , 128, 256]  [512 , 256, 128, 3] 
(d)  [3, 128, 256, 512]  [1024, 512, 256, 3] 
(e)  [3, 16 , 32 , 64 ]  [1024, 512, 256, 3] 
(f)  [3, 128, 256, 512]  [128 , 64 , 32 , 3] 
We then analyzed the importance of size in the discriminative and generative networks. We considered two extreme feature map sizes, where we choose a small and large number of feature maps for the generator and discriminator, and vice versa (see label (e) and (f) in Table 7), and results are shown in Table 8. For LSDCGAN, it can be seen that a large number of feature maps for the discriminator has a better score than a large number of feature maps for the generator. This can also be qualitatively verified by looking at the samples from architectures (a), (e), (f), and (d) in Figure 6. For WDCGAN, we observe the agreement between the LS and IW metric and conflict with MMD and IS. When we look at the samples from the WDCGAN in Figure 5, it is clear that the model with a larger number of feature maps in the discriminator should achieve a better score; this is another example of false intuition propagated by MMD and IS. One interesting observation is that when we compare the score and samples from architecture (a) and (e) from Table 7, architecture (a) is much better than (e) (see Figure 6). This demonstrates that having a large generator and small discriminator is worse than having a small architecture for both networks. Overall, we found that having a larger generator than discriminator does not give good results, and that it is more desirable to have a larger discriminator than generator. Similar results were also observed for MNIST, as shown in S.M. Figure 20. This result somewhat supports the theoretical result from Arora et al. (2017), where the generator capacity needs to be modulated in order for approximately pure equilibrium to exist for GANs.
Model  Architecture  MMD  IW  LS  IS 

(Table 7)  Test vs. Samples  (ResNet)  
WDCGAN  (e)  0.1057 0.0798  450.17 25.74  0.0079 0.0009  6.403 0.839 
(f)  0.2176 0.2706  16.52 15.63  0.0636 0.0101  6.266 0.055  
LSDCGAN  (e)  0.1390 0.1525  343.23 47.55  0.0092 0.0007  5.751 0.511 
(f)  0.0054 0.0022  12.75 4.29  0.0372 0.0068  6.600 0.061 
Lastly, we experimented with how performance changes with respect to the dimension of the noise vectors. The source of the sample starts by transforming a noise vector into a meaningful image. It is unclear how the size of noise affects the ability of the generator to generate a meaningful image. Che et al. (2017) have observed that a 100d noise vector preserves modes better than a 200d noise vector for DCGAN. Our experiments show that this depends on the model. Given a fixed size architecture (d) from Table 7, we observed the performance of LSDCGAN and WDCGAN by varying the size of noise vector . Table 9 illustrates that LSDCGAN gives the best score with a noise dimension of 50 and WDCGAN gives best score with a noise dimension of 150 for both IW and LS. The outcome of LSDCGAN is consistent with the result in (Che et al., 2017). It is possible that this occurs because both models fall into the category of divergences, whereas the WDCGAN behaves differently because its metric falls under a different category, the Integral Probability Metric.
  LSDCGAN  WDCGAN  

IW  LS  IW  LS  
50  3.9010 0.60  0.0547 0.0059  6.0948 3.21  0.0532 0.0069 
100  5.6588 1.47  0.0511 0.0065  5.7358 3.25  0.0528 0.0051 
150  5.8350 0.80  0.0434 0.0036  3.6945 1.33  0.0521 0.0050 
4.3.2 Performance change with respect to the ratio of number of updates between the generator and discriminator
In practice, we alternate between updating the discriminator and generator, and yet this is not guaranteed to give the same result as the solution to the minmax problem in Equation 2. Hence, the update ratio can influence the performance of GANs. We experimented with three different update ratios, , , and , with respect to the discriminator and generator update. We applied these ratios to both the MNIST and CIFAR10 datasets on all models.
Figure 7 presents the LS scores on both MNIST and CIFAR10 and this result is consistent with the IW metric as well (see S.M. Figure 26). However, we did not find that any one update ratio was superior over others between the two datasets. For CIFAR10, the update ratio worked best for all models, and for MNIST, different ratios worked better for different models. Hence, we conclude that number of update ratios for each model needs to be dynamically tuned. The corresponding samples from the models trained by different update ratios are shown in S.M. Figure 27.
4.3.3 Performance with respect to the amount of available training data
In practice, DCGANs are known to be unstable, and the generator tends to suffer as the discriminator improves due to disjoint support between the data and generator distributions (Goodfellow et al., 2014; Arjovsky & Bottou, 2017). WDCGAN and LSDCGAN offer alternative ways to solving this problem. If the model is suffering from disjoint support, having more training examples will not help, and alternatively, if the model does not suffer from such a problem, having more training examples could potentially help.
Here, we explore the sensitivity of three different kinds of GANs with respect to the number of training examples. We have trained GANs with 10,000, 20,000, 30,000, 40,000, and 45,000 examples on CIFAR10. Figure 8 shows that the LS score curve of DCGAN grows quite slowly when compared to WDCGAN and LSDCGAN. The three GANs have a relatively similar loss when they are trained with 10,000 training examples. However, the DCGAN only gained by increasing from 10,000 to 40,000 training examples, whereas the performance of WDCGAN and LSDCGAN improved by and , respectively. Thus, we empirically observe that WDCGAN and LSDCGAN have faster performance increases than a DCGAN as the number of training examples grows.
5 Conclusion
In this paper, we proposed to use four wellknown distance functions as evaluation metrics, and empirically investigated the DCGAN, WDCGAN, and LSDCGAN families under these metrics. Previously, these models were compared based on visual assessment of sample quality and difficulty of training. In our experiments, we showed that there are performance differences in terms of average experiments, but that some are not statistically significant. Moreover, we thoroughly analyzed the performance of GANs under different hyperparameter settings.
There are still several types of GANs that need to be evaluated, such as GRAN (Im et al., 2016a), IWDCGAN (Gulrajani et al., 2017), BEGAN (Berthelot et al., 2017), MMDGAN (Li et al., 2017), and CramerGAN (Bellemare et al., 2017). We hope to evaluate all of these models under this framework and thoroughly analyze them in the future. Moreover, there has been an investigation into taking ensemble approaches to GANs, such as Generative Adversarial Parallelization Im et al. (2016b). Ensemble approaches have been empirically shown to work well in many domains of research, so it would be interesting to find out whether ensembles can also help in minmax problems. Alternatively, we can also try to evaluate other loglikelihoodbased models like NVIL (Mnih & Gregor, 2014), VAE (Kingma & Welling, 2014), DVAE (Im et al., 2015), DRAW (Gregor et al., 2015), RBMs (Hinton et al., 2006; Salakhutdinov & Hinton, 2009), NICE (Dinh et al., 2014), etc.
Model evaluation is an important and complex topic. Model selection, model design, and even research direction can change depending on the evaluation metric. Thus, we need to continuously explore different metrics and rigorously evaluate new models.
References
 (1) pytorchgenerativemodelcollections. https://github.com/znxlwm/pytorchgenerativemodelcollections. Accessed: 20180130.
 Arjovsky & Bottou (2017) Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In arXiv preprint arXiv:1701.04862, 2017.
 Arora et al. (2017) Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Generalization and equilibrium in generative adversarial nets. In arXiv preprint arXiv:1703.00573, 2017.
 Bellemare et al. (2017) Marc G. Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, and Rémi Munos. The cramer distance as a solution to biased wasserstein gradients. In arXiv preprint arXiv:1705.10743, 2017.
 Berthelot et al. (2017) David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. In arXiv preprint arXiv:1703.10717, 2017.
 Che et al. (2017) Tong Che, Yanran Li, Athul Paul Jacob, Yoshua Bengio, and Wenjie Li. Mode regularized generative adversarial networks. In arXiv preprint arXiv:1705.08584, 2017.
 Danihelka et al. (2017) Ivo Danihelka, Balaji Lakshminarayanan, Benigno Uria, Daan Wierstra, and Peter Dayan. Comparison of maximum likelihood and ganbased training of real nvps. In arXiv preprint arXiv:1705.05263, 2017.
 Dinh et al. (2014) Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: nonlinear independent components estimation. In arXiv preprint arXiv:1410.8516, 2014.
 Goodfellow (2016) Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160, 2016.
 Goodfellow et al. (2014) Ian J. Goodfellow, Jean PougetAbadie, Mehdi Mirza, Bing Xu, David WardeFarley, Sherjil Ozair†, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the Neural Information Processing Systems (NIPS), 2014.

Gregor et al. (2015)
Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Jimenez Rezende, and Daan
Wierstra.
Draw: A recurrent neural network for image generation.
InProceedings of the International Conference on Machine Learning (ICML)
, 2015.  Gretton et al. (2012) Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Scholkopf, and Alexander Smola. A kernel twosample test. Journal of Machine Learning Research, 13:723–773, 2012.
 Gulrajani et al. (2017) Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron Courville. Improved training of wasserstein gans. In arXiv preprint arXiv:1704.00028, 2017.
 He et al. (2015) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In arXiv preprint arXiv:1512.03385, 2015.
 Heusel et al. (2017) Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two timescale update rule converge to a local nash equilibrium. In arXiv preprint arXiv:1706.08500, 2017.
 Hinton et al. (2006) Geoffrey E. Hinton, Simon Osindero, and Yee Whye Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18:1527–1554, 2006.

Huang et al. (2017)
Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger.
Densely connected convolutional networks.
In
Proceedings of the Computer Vision and Pattern Recognition (CVPR)
, 2017.  Im et al. (2015) Daniel Jiwoong Im, Sungjin Ahn, Roland Memisevic, and Yoshua Bengio. Denoising criterion for variational autoencoding framework. In arXiv preprint arXiv:1511.06406, 2015.
 Im et al. (2016a) Daniel Jiwoong Im, Dongjoo Kim, Hui Jiang, and Roland Memisevic. Generating images with recurrent adversarial networks. In arXiv preprint arXiv:1602.05110, 2016a.
 Im et al. (2016b) Daniel Jiwoong Im, He Ma, Dongjoo Kim, and Graham Taylor. Generative adversarial paralleliation. In arXiv preprint arXiv:1612.04021, 2016b.
 Im et al. (2016c) Daniel Jiwoong Im, Michael Tao, and Kristin Branson. An empirical analysis of the optimization of deep network loss surfaces. In arXiv preprint arXiv:1612.04010, 2016c.
 Junbo Zhao (2016) Yann LeCun Junbo Zhao, Michael Mathieu. Energybased generative adversarial network. In arXiv preprint arXiv:1609.03126, 2016.
 Kingma & Welling (2014) Diederik P Kingma and Max Welling. Autoencoding varational bayes. In Proceedings of the Neural Information Processing Systems (NIPS), 2014.
 Kodali et al. (2017) Naveen Kodali, Jacob Abernethy, James Hays, and Zsolt Kira. On convergence and stability of gans. In arXiv preprint arXiv:1705.07215, 2017.

Krizhevsky et al. (2012)
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton.
Imagenet classification with deep convolutional neural networks.
In Proceedings of the Neural Information Processing Systems (NIPS), 2012.  LeCun et al. (1998) Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradientbased learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
 Li et al. (2017) ChunLiang Li, WeiCheng Chang, Yu Cheng, Yiming Yang, and Barnabás Póczos. Mmd gan: Towards deeper understanding of moment matching network. In arXiv preprint arXiv:1705.08584, 2017.
 LopezPaz & Oquab (2016) David LopezPaz and Maxime Oquab. Revisiting classifier twosample tests. In arXiv preprint arXiv:1610.06545, 2016.
 Mao et al. (2017) Xudong Mao, Qing Li, Haoran Xie, Raymond Y.K. Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In arXiv preprint arXiv:1611.04076, 2017.
 Mescheder et al. (2017) Lars Mescheder, Sebastian Nowozin, and Andreas Geiger. The numerics of gans. In Proceedings of the Neural Information Processing Systems (NIPS), 2017.
 Mnih & Gregor (2014) Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. In Proceedings of the International Conference on Machine Learning (ICML), 2014.
 Muller (1997) Alfred Muller. Integral probability metrics and their generating classes of functions. Advances in Applied Probability, 29(2):429–443, 1997.
 Nguyen et al. (2008) XuanLong Nguyen, Martin J. Wainwright, and Michael I. Jordan. Estimating divergence functionals and the likelihood ratio by penalized convex risk minimization. In Proceedings of the Neural Information Processing Systems (NIPS), 2008.
 Nowozin et al. (2016) Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. fgan: Training generative neural samplers using variational divergence minimization. In arXiv preprint arXiv:1606.00709, 2016.
 Radford et al. (2015) Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In arXiv preprint arXiv:1511.06434, 2015.

Salakhutdinov & Hinton (2009)
Ruslan Salakhutdinov and Geoffrey E. Hinton.
Deep boltzmann machines.
In Proceedings of the International Conference on Machine Learning (ICML), 2009.  Salimans et al. (2016) Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In arXiv preprint arXiv:1606.03498, 2016.
 Simonyan & Zisserman (2014) Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for largescale image recognition. In arXiv preprint arXiv:1409.1556, 2014.
 Sriperumbudur et al. (2009) Bharath Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Bernhard Schölkopf, and Gert Lanckriet. On integral probability metrics, phidivergences and binary classification. 01 2009.
 Sutherland et al. (2017) Dougal J. Sutherland, HsiaoYu Tung, Heiko Strathmann, Soumyajit De Aaditya Ramdas, Alex Smola, and Arthur Gretton. Generative models and model criticism via optimized maximum mean discrepancy. In arXiv preprint arXiv:1611.04488, 2017.
 Szegedy et al. (2015) Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In arXiv preprint arXiv:1512.00567, 2015.
 Theis et al. (2015) Lucas Theis, Aäron van den Oord, and Matthias Bethge. A note on the evaluation of generative models. 5 November 2015.
 Villani (2009) Cedric Villani. Grundlehren der mathematischen wissenschaften. In Optimal Transport: Old and New. Springer, Berline, 2009.
 Wu et al. (2016) Yuhuai Wu, Yuri Burda, Ruslan Salakhutdinov, and Roger Grosse. On the quantitative analysis of decoder based generative models. In International Conference on Learning Representation, 2016.
Appendix
Relationship between metrics and binary classification
In this paper, we considered four distance metrics that belong to two class of metrics, divergence and IPMs. Sriperumbudur et al. (2009) have shown that the optimal risk function is associated with a binary classifier with and distributions conditioned on a class when the discriminant function is restricted to certain (Theorem 17 from (Sriperumbudur et al., 2009)).
Let the optimal risk function be:
(11) 
where is the set of discriminant functions (classifier), , and is the loss function.
By following derivation, we can see that the optimal risk function becomes IPM:
(12)  
(13)  
(14)  
(15) 
where and .
The second equality is derived by separating the loss for class 1 and class 0. The third equality is from the way how we chose L(1,f(x)) and L(0,f(x)). The last equality is derived from that fact that is symmetric around zero . Hence, this shows that with appropriately choosing , MMD and Wasserstein distance can be understood as the optimal risk associated with binary classifier with specific set of functions. For example, Wasserstein distance and MMD distances are equivalent to the optimal risk function with 1Lipschitz classifiers and a RKHS classifier with an unit length.
Experimental Hyperparameters
GAN training  Critic Training (test time)  
Model  Disc. Lr.  Gen. Lr.  Ratio^{3}^{3}3Number of updates ratio between discriminator and generator.  Cr. Lr.  Cr. Kern  Num Epoch  
Table 5  DCGAN  0.0002  0.0004  1:2  0.0001  [1, 128, 32]  25 
WDCGAN  0.0004  0.0008  1:1  
LSDCGAN  0.0004  0.0008  1:2  
Table 5  DCGAN  0.0002  0.0001  1:2  0.0002  [3, 128, 256, 512]  11 
WDCGAN  0.0008  0.0004  1:1  
LSDCGAN  0.0008  0.0004  1:2  
Table 5  DCGAN  0.00005  0.0001  1:2  0.0002  [3, 128, 256, 512,1024]  4 
WDCGAN  0.0002  0.0004  1:2  
LSDCGAN  0.0002  0.0004  1:2  
Table 5  ALL GANs  0.0002  0.0002  1:1  0.0002  [1, 64, 128]  25 
Table 8  DCGAN  0.0002  0.0001  1:2  0.0002  [3, 128, 256, 512]  11 
WDCGAN  0.0002  0.0001  1:1  
LSDCGAN  0.0008  0.0004  1:2  
Table 12  ALL GANs  0.0002  0.0002  1:1  0.0002  [1, 64, 128]  25 
Table 12  ALL GANs  0.0002  0.0002  1:1  0.0002  [1, 64, 128]  25 
Figure 7  DCGAN  0.0001  0.00005  5:1  0.0002  [3, 128, 256, 512]  11 
1:1  
1:5  
WDCGAN  0.0008  0.0004  5:1  
1:1  
1:5  
LSDCGAN  5:1  
1:1  
1:5  
Figure 26  DCGAN  0.0001  0.00005  5:1  0.0002  [1, 128, 32]  25 
1:1  
1:5  
WDCGAN  0.0008  0.0004  5:1  
1:1  
1:5  
LSDCGAN  5:1  
1:1  
1:5  
Figure 17  DCGAN  0.0002  0.0001  1:2  0.0002  [3, 128, 256, 512]  11 
WDCGAN  0.0002  0.0001  1:1  
LSDCGAN  0.0008  0.0004  1:2  
Figure 29  DCGAN  0.0002  0.0001  1:5  0.0002  [3, 256, 512, 1028]  11 
WDCGAN  0.0008  0.0004  1:1  
LSDCGAN  0.0008  0.0004  1:5 
More Experiments
Model  LS Score  IW Score  

Trained on training data  Trained on validation. data  Trained on training data  Trained on validation. data  
DCGAN  0.312 0.010  0.4408 0.0201  0.300 0.0103  0.259 0.0083 
EBGAN  3.38e6 0.1.86e7  3.82e6 2.82e7  0.999 0.0001  0.999 0.0001 
WGAN GP  0.196 0.006  0.307 0.0381  0.705 0.0202  0.635 0.0270 
LSGAN  0.323 0.0104  0.352 0.0143  0.232 0.0156  0.195 0.0103 
BEGAN  0.081 0.016  0.140 0.0329  0.888 0.0097  0.858 0.0131 
DRAGAN  0.318 0.012  0.384 0.0139  0.266 0.0060  0.235 0.0079 
Model  LS Score  IW Score  

Trained on training data  Trained on validation. data  Trained on training data  Trained on validation. data  
DCGAN  0.1638 0.010  0.1635 0.0006  0.408 0.0135  0.4118 0.0107 
EBGAN  0.0037 0.0009  0.0048 0.0023  0.415 0.0067  0.4247 0.0098 
WGAN GP  0.000175 0.0000876  0.000448 0.0000862  0.921 0.0061  0.9234 0.0059 
LSGAN  0.135 0.0046  0.136 0.0074  0.631 0.0106  0.6236 0.0200 
BEGAN  0.1133 0.042  0.0893 0.0095  0.429 0.0148  0.4293 0.0213 
DRAGAN  0.1638 0.015  0.1645 0.0151  0.641 0.0304  0.6311 0.0547 
We trained two critics on training data and validation data, respectively, and evaluated on test data from both critics. We trained six GANs (GAN, LSDCGAN, WDCGAN GP, DRAGAN, BEGAN, EBGAN) on MNIST and FashionMNIST. We trained these GANs with 50,000 training examples. At test time, we used 10,000 training and 10,000 validation examples for training the critics, and evaluated on 10,000 test examples. Here, we present the test scores from the critics trained on training and validation data. The results are shown in Table 12 and 12. Note that we also have the IW and FID evaluation on these models in the paper. For FashionMNIST, we find that test scores with a critic trained on training and validation data are very close. Hence, we do not see any indication of overfitting. On the other hand, there are gaps between the scores for the MNIST dataset and the test scores from critics trained on the validation set. which gives better performance than the ones that are trained on the training set.
Ratio  DCGAN Samples 

1:1  
1:5 
Ratio  WDCGAN Samples 

5:1  
1:1  
1:5 
Ratio  LSDCGAN Samples 

5:1  
1:1  
1:5 