{"title": "Improved Techniques for Training GANs", "book": "Advances in Neural Information Processing Systems", "page_first": 2234, "page_last": 2242, "abstract": "We present a variety of new architectural features and training procedures that we apply to the generative adversarial networks (GANs) framework. Using our new techniques, we achieve state-of-the-art results in semi-supervised classification on MNIST, CIFAR-10 and SVHN. The generated images are of high quality as confirmed by a visual Turing test: Our model generates MNIST samples that humans cannot distinguish from real data, and CIFAR-10 samples that yield a human error rate of 21.3%. We also present ImageNet samples with unprecedented resolution and show that our methods enable the model to learn recognizable features of ImageNet classes.", "full_text": "Improved Techniques for Training GANs\n\nTim Salimans\n\ntim@openai.com\n\nIan Goodfellow\nian@openai.com\n\nWojciech Zaremba\nwoj@openai.com\n\nVicki Cheung\n\nvicki@openai.com\n\nAlec Radford\n\nalec@openai.com\n\nXi Chen\n\npeter@openai.com\n\nAbstract\n\nWe present a variety of new architectural features and training procedures that we\napply to the generative adversarial networks (GANs) framework. Using our new\ntechniques, we achieve state-of-the-art results in semi-supervised classi\ufb01cation on\nMNIST, CIFAR-10 and SVHN. The generated images are of high quality as con-\n\ufb01rmed by a visual Turing test: our model generates MNIST samples that humans\ncannot distinguish from real data, and CIFAR-10 samples that yield a human error\nrate of 21.3%. We also present ImageNet samples with unprecedented resolution\nand show that our methods enable the model to learn recognizable features of\nImageNet classes.\n\n1\n\nIntroduction\n\nGenerative adversarial networks [1] (GANs) are a class of methods for learning generative models\nbased on game theory. The goal of GANs is to train a generator network G(z; \u03b8(G)) that produces\nsamples from the data distribution, pdata(x), by transforming vectors of noise z as x = G(z; \u03b8(G)).\nThe training signal for G is provided by a discriminator network D(x) that is trained to distinguish\nsamples from the generator distribution pmodel(x) from real data. The generator network G in turn\nis then trained to fool the discriminator into accepting its outputs as being real.\nRecent applications of GANs have shown that they can produce excellent samples [2, 3]. However,\ntraining GANs requires \ufb01nding a Nash equilibrium of a non-convex game with continuous, high-\ndimensional parameters. GANs are typically trained using gradient descent techniques that are\ndesigned to \ufb01nd a low value of a cost function, rather than to \ufb01nd the Nash equilibrium of a game.\nWhen used to seek for a Nash equilibrium, these algorithms may fail to converge [4].\nIn this work, we introduce several techniques intended to encourage convergence of the GANs game.\nThese techniques are motivated by a heuristic understanding of the non-convergence problem. They\nlead to improved semi-supervised learning peformance and improved sample generation. We hope\nthat some of them may form the basis for future work, providing formal guarantees of convergence.\nAll code and hyperparameters may be found at https://github.com/openai/improved-gan.\n\n2 Related work\n\nSeveral recent papers focus on improving the stability of training and the resulting perceptual quality\nof GAN samples [2, 3, 5, 6]. We build on some of these techniques in this work. For instance, we\nuse some of the \u201cDCGAN\u201d architectural innovations proposed in Radford et al. [3], as discussed\nbelow.\nOne of our proposed techniques, feature matching, discussed in Sec. 3.1, is similar in spirit to\napproaches that use maximum mean discrepancy [7, 8, 9] to train generator networks [10, 11].\n\n30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.\n\n\fAnother of our proposed techniques, minibatch features, is based in part on ideas used for batch\nnormalization [12], while our proposed virtual batch normalization is a direct extension of batch\nnormalization.\nOne of the primary goals of this work is to improve the effectiveness of generative adversarial\nnetworks for semi-supervised learning (improving the performance of a supervised task, in this case,\nclassi\ufb01cation, by learning on additional unlabeled examples). Like many deep generative models,\nGANs have previously been applied to semi-supervised learning [13, 14], and our work can be seen\nas a continuation and re\ufb01nement of this effort. In concurrent work, Odena [15] proposes to extend\nGANs to predict image labels like we do in Section 5, but without our feature matching extension\n(Section 3.1) which we found to be critical for obtaining state-of-the-art performance.\n\n3 Toward Convergent GAN Training\n\nTraining GANs consists in \ufb01nding a Nash equilibrium to a two-player non-cooperative game.\nEach player wishes to minimize its own cost function, J (D)(\u03b8(D), \u03b8(G)) for the discriminator and\nJ (G)(\u03b8(D), \u03b8(G)) for the generator. A Nash equilibirum is a point (\u03b8(D), \u03b8(G)) such that J (D) is at a\nminimum with respect to \u03b8(D) and J (G) is at a minimum with respect to \u03b8(G). Unfortunately, \ufb01nd-\ning Nash equilibria is a very dif\ufb01cult problem. Algorithms exist for specialized cases, but we are not\naware of any that are feasible to apply to the GAN game, where the cost functions are non-convex,\nthe parameters are continuous, and the parameter space is extremely high-dimensional.\nThe idea that a Nash equilibrium occurs when each player has minimal cost seems to intuitively mo-\ntivate the idea of using traditional gradient-based minimization techniques to minimize each player\u2019s\ncost simultaneously. Unfortunately, a modi\ufb01cation to \u03b8(D) that reduces J (D) can increase J (G), and\na modi\ufb01cation to \u03b8(G) that reduces J (G) can increase J (D). Gradient descent thus fails to converge\nfor many games. For example, when one player minimizes xy with respect to x and another player\nminimizes \u2212xy with respect to y, gradient descent enters a stable orbit, rather than converging to\nx = y = 0, the desired equilibrium point [16]. Previous approaches to GAN training have thus\napplied gradient descent on each player\u2019s cost simultaneously, despite the lack of guarantee that this\nprocedure will converge. We introduce the following techniques that are heuristically motivated to\nencourage convergence:\n\n3.1 Feature matching\nFeature matching addresses the instability of GANs by specifying a new objective for the generator\nthat prevents it from overtraining on the current discriminator. Instead of directly maximizing the\noutput of the discriminator, the new objective requires the generator to generate data that matches\nthe statistics of the real data, where we use the discriminator only to specify the statistics that we\nthink are worth matching. Speci\ufb01cally, we train the generator to match the expected value of the\nfeatures on an intermediate layer of the discriminator. This is a natural choice of statistics for the\ngenerator to match, since by training the discriminator we ask it to \ufb01nd those features that are most\ndiscriminative of real data versus data generated by the current model.\nLetting f (x) denote activations on an intermediate layer of the discriminator, our new objective for\nthe generator is de\ufb01ned as: ||Ex\u223cpdataf (x) \u2212 Ez\u223cpz(z)f (G(z))||2\n2. The discriminator, and hence\nf (x), are trained in the usual way. As with regular GAN training, the objective has a \ufb01xed point\nwhere G exactly matches the distribution of training data. We have no guarantee of reaching this\n\ufb01xed point in practice, but our empirical results indicate that feature matching is indeed effective in\nsituations where regular GAN becomes unstable.\n\n3.2 Minibatch discrimination\nOne of the main failure modes for GAN is for the generator to collapse to a parameter setting where\nit always emits the same point. When collapse to a single mode is imminent, the gradient of the\ndiscriminator may point in similar directions for many similar points. Because the discriminator\nprocesses each example independently, there is no coordination between its gradients, and thus no\nmechanism to tell the outputs of the generator to become more dissimilar to each other. Instead,\nall outputs race toward a single point that the discriminator currently believes is highly realistic.\nAfter collapse has occurred, the discriminator learns that this single point comes from the generator,\nbut gradient descent is unable to separate the identical outputs. The gradients of the discriminator\n\n2\n\n\fthen push the single point produced by the generator around space forever, and the algorithm cannot\nconverge to a distribution with the correct amount of entropy. An obvious strategy to avoid this type\nof failure is to allow the discriminator to look at multiple data examples in combination, and perform\nwhat we call minibatch discrimination.\nThe concept of minibatch discrimination is quite general: any discriminator model that looks\nat multiple examples in combination, rather than in isolation, could potentially help avoid col-\nlapse of the generator.\nIn fact, the successful application of batch normalization in the dis-\ncriminator by Radford et al. [3] is well explained from this perspective. So far, however, we\nhave restricted our experiments to models that explicitly aim to identify generator samples that\nare particularly close together. One successful speci\ufb01cation for modelling the closeness between\nexamples in a minibatch is as follows: Let f (xi) \u2208 RA denote a vector of features for in-\nput xi, produced by some intermediate layer in the discriminator. We then multiply the vector\nf (xi) by a tensor T \u2208 RA\u00d7B\u00d7C, which results in a matrix Mi \u2208 RB\u00d7C. We then compute\nthe L1-distance between the rows of the resulting matrix Mi across samples i \u2208 {1, 2, . . . , n}\ncb(xi, xj) = exp(\u2212||Mi,b \u2212 Mj,b||L1) \u2208 R.\nand apply a negative exponential (Fig. 1):\nThe output o(xi) for this minibatch layer for a sample xi\nis then de\ufb01ned as the sum of the cb(xi, xj)\u2019s to all other\nsamples:\n\no(xi)b =\n\ncb(xi, xj) \u2208 R\n\nn(cid:88)\n(cid:104)\n\nj=1\n\n(cid:105) \u2208 RB\n\no(xi)1, o(xi)2, . . . , o(xi)B\n\no(xi) =\no(X) \u2208 Rn\u00d7B\n\nNext, we concatenate the output o(xi) of the minibatch\nlayer with the intermediate features f (xi) that were its\ninput, and we feed the result into the next layer of the\ndiscriminator. We compute these minibatch features sep-\narately for samples from the generator and from the train-\ning data. As before, the discriminator is still required to\noutput a single number for each example indicating how\nlikely it is to come from the training data: The task of the discriminator is thus effectively still to\nclassify single examples as real data or generated data, but it is now able to use the other examples in\nthe minibatch as side information. Minibatch discrimination allows us to generate visually appealing\nsamples very quickly, and in this regard it is superior to feature matching (Section 6). Interestingly,\nhowever, feature matching was found to work much better if the goal is to obtain a strong classi\ufb01er\nusing the approach to semi-supervised learning described in Section 5.\n\nFigure 1: Figure sketches how mini-\nbatch discrimination works. Features\nf (xi) from sample xi are multiplied\nthrough a tensor T , and cross-sample\ndistance is computed.\n\n(cid:80)t\n3.3 Historical averaging\nWhen applying this technique, we modify each player\u2019s cost to include a term ||\u03b8 \u2212 1\ni=1 \u03b8[i]||2,\nwhere \u03b8[i] is the value of the parameters at past time i. The historical average of the parameters can\nbe updated in an online fashion so this learning rule scales well to long time series. This approach is\nloosely inspired by the \ufb01ctitious play [17] algorithm that can \ufb01nd equilibria in other kinds of games.\nWe found that our approach was able to \ufb01nd equilibria of low-dimensional, continuous non-convex\ngames, such as the minimax game with one player controlling x, the other player controlling y, and\nvalue function (f (x) \u2212 1)(y \u2212 1), where f (x) = x for x < 0 and f (x) = x2 otherwise. For\nthese same toy games, gradient descent fails by going into extended orbits that do not approach the\nequilibrium point.\n\nt\n\n3.4 One-sided label smoothing\nLabel smoothing, a technique from the 1980s recently independently re-discovered by Szegedy et.\nal [18], replaces the 0 and 1 targets for a classi\ufb01er with smoothed values, like .9 or .1, and was\nrecently shown to reduce the vulnerability of neural networks to adversarial examples [19].\nReplacing positive classi\ufb01cation targets with \u03b1 and negative targets with \u03b2, the optimal discriminator\nbecomes D(x) = \u03b1pdata(x)+\u03b2pmodel(x)\npdata(x)+pmodel(x) . The presence of pmodel in the numerator is problematic\nbecause, in areas where pdata is approximately zero and pmodel is large, erroneous samples from\n\n3\n\n\fpmodel have no incentive to move nearer to the data. We therefore smooth only the positive labels to\n\u03b1, leaving negative labels set to 0.\n\n3.5 Virtual batch normalization\nBatch normalization greatly improves optimization of neural networks, and was shown to be highly\neffective for DCGANs [3]. However, it causes the output of a neural network for an input example\nx to be highly dependent on several other inputs x(cid:48) in the same minibatch. To avoid this problem\nwe introduce virtual batch normalization (VBN), in which each example x is normalized based on\nthe statistics collected on a reference batch of examples that are chosen once and \ufb01xed at the start\nof training, and on x itself. The reference batch is normalized using only its own statistics. VBN is\ncomputationally expensive because it requires running forward propagation on two minibatches of\ndata, so we use it only in the generator network.\n\n4 Assessment of image quality\n\nto\nGenerative adversarial networks lack an objective function, which makes it dif\ufb01cult\ncompare performance of different models.\nOne intuitive metric of performance can be\nobtained by having human annotators judge the visual quality of samples [2]. We\nautomate this process using Amazon Mechanical Turk (MTurk),\nusing the web in-\nterface in \ufb01gure Fig. 2 (live at http://infinite-chamber-35121.herokuapp.com/\ncifar-minibatch/), which we use to ask annotators to distinguish between generated data\nand real data. The resulting quality assessments of our models are described in Section 6.\n\nFigure 2: Web interface given to anno-\ntators. Annotators are asked to distin-\nguish computer generated images from\nreal ones.\n\nA downside of using human annotators is that the metric\nvaries depending on the setup of the task and the moti-\nvation of the annotators. We also \ufb01nd that results change\ndrastically when we give annotators feedback about their\nmistakes: By learning from such feedback, annotators are\nbetter able to point out the \ufb02aws in generated images, giv-\ning a more pessimistic quality assessment. The left col-\numn of Fig. 2 presents a screen from the annotation pro-\ncess, while the right column shows how we inform anno-\ntators about their mistakes.\n\nthe model to generate varied images, so the marginal(cid:82) p(y|x = G(z))dz should have high entropy.\n\nAs an alternative to human annotators, we propose an automatic method to evaluate samples, which\nwe \ufb01nd to correlate well with human evaluation: We apply the Inception model1 [20] to every\ngenerated image to get the conditional label distribution p(y|x). Images that contain meaningful\nobjects should have a conditional label distribution p(y|x) with low entropy. Moreover, we expect\nCombining these two requirements, the metric that we propose is: exp(ExKL(p(y|x)||p(y))), where\nwe exponentiate results so the values are easier to compare. Our Inception score is closely related\nto the objective used for training generative models in CatGAN [14]: Although we had less success\nusing such an objective for training, we \ufb01nd it is a good metric for evaluation that correlates very\nwell with human judgment. We \ufb01nd that it\u2019s important to evaluate the metric on a large enough\nnumber of samples (i.e. 50k) as part of this metric measures diversity.\n\n5 Semi-supervised learning\n\nConsider a standard classi\ufb01er for classifying a data point x into one of K possible classes. Such\na model takes in x as input and outputs a K-dimensional vector of logits {l1, . . . , lK}, that can\n(cid:80)K\nbe turned into class probabilities by applying the softmax: pmodel(y = j|x) =\nk=1 exp(lk). In\nsupervised learning, such a model is then trained by minimizing the cross-entropy between the\nobserved labels and the model predictive distribution pmodel(y|x).\n\nexp(lj )\n\n1We use the pretrained Inception model from http://download.tensorflow.org/models/image/\nimagenet/inception-2015-12-05.tgz. Code to compute the Inception score with this model will be made\navailable by the time of publication.\n\n4\n\n\fWe can do semi-supervised learning with any standard classi\ufb01er by simply adding samples from\nthe GAN generator G to our data set, labeling them with a new \u201cgenerated\u201d class y = K + 1, and\ncorrespondingly increasing the dimension of our classi\ufb01er output from K to K + 1. We may then\nuse pmodel(y = K + 1 | x) to supply the probability that x is fake, corresponding to 1 \u2212 D(x) in\nthe original GAN framework. We can now also learn from unlabeled data, as long as we know that\nit corresponds to one of the K classes of real data by maximizing log pmodel(y \u2208 {1, . . . , K}|x).\nAssuming half of our data set consists of real data and half of it is generated (this is arbitrary), our\nloss function for training the classi\ufb01er then becomes\n\nL = \u2212Ex,y\u223cpdata(x,y)[log pmodel(y|x)] \u2212 Ex\u223cG[log pmodel(y = K + 1|x)]\n= Lsupervised + Lunsupervised, where\n\nLsupervised = \u2212Ex,y\u223cpdata(x,y) log pmodel(y|x, y < K + 1)\nLunsupervised = \u2212{Ex\u223cpdata(x) log[1 \u2212 pmodel(y = K + 1|x)] + Ex\u223cG log[pmodel(y = K + 1|x)]},\n\nwhere we have decomposed the total cross-entropy loss into our standard supervised loss function\nLsupervised (the negative log probability of the label, given that the data is real) and an unsupervised\nloss Lunsupervised which is in fact the standard GAN game-value as becomes evident when we substi-\ntute D(x) = 1 \u2212 pmodel(y = K + 1|x) into the expression:\n\nLunsupervised = \u2212{Ex\u223cpdata(x) log D(x) + Ez\u223cnoise log(1 \u2212 D(G(z)))}.\n\nis\n\nto\n\noptimal\n\nsolution\n\nfor minimizing\n\nboth Lsupervised\n\nand Lunsupervised\n\nThe\nhave\nexp[lj(x)] = c(x)p(y=j, x)\u2200j