Batch Normalization (BN) is a common technique used both in discriminative and generative networks in order to speed-up training. On the other hand, the learnable parameters of BN are commonly used in conditional Generative Adversarial Networks for representing class-specific information using conditional Batch Normalization (cBN). In this paper we propose to generalize both BN and cBN using a Whitening and Coloring based batch normalization. We apply our method to conditional and unconditional image generation tasks and we show that replacing the BN feature standardization and scaling with our feature whitening and coloring improves the final qualitative results and the training speed. We test our approach on different datasets and we show a consistent improvement orthogonal to different GAN frameworks. Our CIFAR-10 supervised results are higher than all previous works on this dataset.READ FULL TEXT VIEW PDF
Generative adversarial networks (GANs) are highly effective unsupervised...
Batch normalization has been widely used to improve optimization in deep...
Batch normalization was introduced in 2015 to speed up training of deep
Generative Adversarial Networks (GANs) are known to be difficult to trai...
Batch Normalization (BN) is extensively employed in various network
Traditional convolution-based generative adversarial networks synthesize...
Generative Networks have shown great promise in generating photo-realist...
Generative Adversarial Networks (GANs) (Goodfellow et al. (2014)) have drawn much attention in the last years due to their proven ability to generate sufficiently realistic short videos (Vondrick et al. (2016)) or still images (Gulrajani et al. (2017)), possibly conditioned on some input information such as: a class label (Odena et al. (2017)), an image (Isola et al. (2017); Tang et al. (2019)), a textual description (Reed et al. (2016)), some structured data (Tang et al. (2018); Siarohin et al. (2018a)) or some perceptual image attribute (Siarohin et al. (2018b))). When the generation process depends on some input data, the framework is commonly called conditional GAN (cGAN).
In this paper we deal with both conditional and unconditional GANs and we propose to replace Batch Normalization (Ioffe and Szegedy (2015)) (BN) with a Whitening and Coloring (WC) transform (Hossain (2016)) in order to jointly speed-up training and increase the network representation capacity. BN, proposed in (Ioffe and Szegedy (2015)) for discriminative tasks and then adopted in many discriminative and generative networks, is based on feature standardization of all the network layers. The original motivation behind the introduction of BN is to alleviate the internal Covariate Shift problem caused by the continuous modification of each layer’s input representation during training (Ioffe and Szegedy (2015)). However, recent works (Santurkar et al. (2018); Kohler et al. (2018)) showed that a major motivation of the empirical success of BN relies also on a higher training stability, the latter being due to a better-conditioning of the Jacobian of the corresponding loss when BN is used compared to training without BN. Basically, feature normalization makes the loss landscape smoother, hence it stabilizes and speeds-up training. Based on this intuition, BN has very recently been extended in (Huang et al. (2018)) to full-feature decorrelation using ZCA whitening (Kessy et al. (2017)) in a discriminative scenario.
On the other hand, GAN training is known to be particularly unstable, because GAN optimization aims at finding a Nash equilibrium between two players, a problem which is more difficult and less stable than the common discriminative-network optimization and that frequently leads to non-convergence. Specifically, Odena et al. (2018) show the relation between stability of GAN training and conditioning of the Jacobian. This motivates our proposal to extend BN to full-feature whitening in a GAN scenario (more details on this in Sec. 6). Differently from Huang et al. (2018), our feature whitening is based on the Cholesky decomposition (Dereniowski and Marek (2004)), and we empirically show that our method is much faster and stable and achieves better results in GAN training with respect to ZCA whitening. Importantly, while in (Huang et al. (2018)) feature whitening is followed by a per-dimension scaling and shifting of each feature, as in the original BN (Ioffe and Szegedy (2015)), we propose to use a coloring transform (Hossain (2016)
) to keep the network representation capacity unchanged with respect to a non-whitened layer. Our coloring projects each whitened feature vector in a new distribution using learnablemultivariate filters and we show that this transformation is critical when using whitening for GANs.
The second contribution of this paper is based on the aforementioned coloring transform to represent class-specific information in a cGAN scenario. Formally, given a set of class labels , we want to generate an image , where the generator is input with a noise vector and a class label . For instance, if , we want a foreground object in depicting a cat, if , then should represent a dog, etc. The categorical conditional information is commonly represented in using conditional batch normalization (cBN) (Dumoulin et al. (2016b); Gulrajani et al. (2017); Miyato and Koyama (2018)), where a class-specific pair of scaling-shifting parameters is learned for each class . In our proposed conditional Whitening and Coloring (cWC) the class-specific scaling-shifting parameters are replaced with a combination of class-agnostic and class-specific coloring filters. The intuitive idea behind the proposed cWC is that multivariate transformation filters are more informative than the scalar parameters proposed in (Dumoulin et al. (2016b)) and, thus, they can more accurately represent class-specific information.
Finally, in order to have a set of class-specific coloring filters () which grows sub-linearly with respect to the number of classes , we represent each using a common, restricted dictionary of class-independent filters together with class-filter soft-assignment weights. The use of a class-independent dictionary of filters, together with a class-agnostic branch in the architecture of our cGAN generators (see Sec. 4) makes it possible to share parameters over all the classes and to reduce the final number of weights which need to be learned.
In summary, our contributions are the following:
We propose a whitening-coloring transform for batch normalization to improve stability of GAN training which is based on the Cholesky decomposition and a re-projection of the whitened features in a learned distribution.
In a cGAN framework we propose: (1) A class-specific coloring transform which generalizes cBN (Dumoulin et al. (2016b)) and (2) A soft-assignment of classes to filters to reduce the network complexity.
Using different datasets and basic GAN-cGAN frameworks, we show that our WC and cWC consistently improve the original frameworks’ results and training speed.
Our code is publicly available https://github.com/AliaksandrSiarohin/wc-gan.
Batch Normalization. BN (Ioffe and Szegedy (2015)) is based on a batch-based (per-dimension) standardization of all the network’s layers. Formally, given a batch of -dimensional samples , input to a given network layer, BN transforms each sample in using:
where: () indicates the -th dimension of the data, and
are, respectively, the mean and the standard deviation computed with respect to the-th dimension of the samples in and is a constant used to prevent numerical instability. Finally, and are learnable scaling and shifting parameters, which are introduced in (Ioffe and Szegedy (2015)) in order to prevent losing the original representation capacity of the network. More in detail, and are used to guarantee that the BN transform can be inverted, if necessary, in order to represent the identity transform (Ioffe and Szegedy (2015)). BN is commonly used in both discriminative and generative networks and has been extended in many directions. For instance, Ba et al. (2016) propose Layer Normalization, where normalization is performed using all the layer’s activation values of a single sample. Luo (2017) proposes to whiten the activation values using a learnable whitening matrix. Differently from (Luo (2017)), our whitening matrix only depends on the specific batch of data. Similarly to our work, batch-dependent full-feature decorrelation is performed also by Huang et al. (2018), who use a ZCA-whitening and show promising results in a discriminative scenario. However, the whitening part of our method is based on the Cholesky decomposition and we empirically show that, in a GAN scenario where training is known to be highly unstable, our method produces better results and it is much faster to compute. Moreover, while in (Huang et al. (2018)) only scaling and shifting parameters are used after feature decorrelation, we propose to use a coloring transform. Analogously to the role played by and in Eq. 1, coloring can potentially invert the whitening transform if that were the optimal thing to do, in this way preserving the original representation capacity of the network . We empirically show (see Sec. 5.1.1) that whitening without coloring performs poorly. Last but not least, our conditional coloring can be used to represent rich class-specific information in a cGAN framework (see below).
Representing conditional information in cGANs. cGANs have been widely used to produce images conditioned on some input. For instance, Isola et al. (2017) transform a given image in a second image represented in another “channel”. Siarohin et al. (2018a) extend this framework to deal with structured conditional information (the human body pose). In (Reed et al. (2016)) a textual description of an image is used to condition the image generation. When the conditioning data is a categorical, unstructured variable, as in the case investigated in this paper, information about can be represented in different ways. cBN (Dumoulin et al. (2016b)) uses to select class-specific scaling and shifting parameters in the generator (more details in Sec. 4). Another common approach is to represent using a one-hot vector, which is concatenated with the input layer of both the discriminator and the generator (Mirza and Osindero (2014); Perarnau et al. (2016)) or with the first convolutional layer of the discriminator (Perarnau et al. (2016)). Miyato and Koyama (2018)
split the last discriminator layer in two branches. The first branch estimates the class-agnostic probability that the input image is real, while the second branch represents class-conditional information. This solution is orthogonal to our proposed cWC and we show in Sec.5.2 that the two approaches can be combined for boosting the results obtained in (Miyato and Koyama (2018)).
We start describing the details of our Whitening and Coloring transform in the unconditional case, which we refer to as WC in the rest of the paper. We always assume an image domain and convolutional generator/discriminator networks. However, most of the solutions we propose can potentially be applied to non-visual domains and non-convolutional networks.
be the tensor representing the activation values of the convolutional feature maps for a given image and layer, withchannels and spatial locations. Similarly to BN, we consider the -dimensional activations of each location in the convolutional grid as a separate instance in . This is done also to have a larger cardinality batch which alleviates instability issues when computing the batch-related statistics. In fact, a mini-batch of images corresponds to (Ioffe and Szegedy (2015)).
Using a vectorial notation, our proposed WC-based batch normalization is given by , where :
In Eq. 3, the vector is the mean of the elements in (being in Eq. 1 its -th component), while the matrix is such that: , where is the covariance matrix computed using . This transformation performs the full whitening of and the resulting set of vectors
lies in a spherical distribution (i.e., with a covariance matrix equal to the identity matrix). Note that, similarly toand in Eq. 1, and are completely data-dependent, being computed using only and without any learnable parameter involved. On the other hand, Eq. 2 performs the coloring transform (Hossain (2016)), which projects the elements in
onto a multivariate Gaussian distribution with an arbitrary covariance matrix (see Sec.2 for the motivation behind this projection). Eq. 2 is based on the learnable parameters and , where is a vector of parameters (a coloring filter) which generalizes the scalar parameter in Eq. 1. Coloring is a linear operation and can be simply implemented using convolutional filters and the corresponding biases . We compute using the Cholesky decomposition, which has the important properties to be efficient and well conditioned. We provide the details in Sec. A, where we also analyze the computational costs, while in Sec. E we compare our approach with the ZCA-based whitening proposed by Huang et al. (2018).
WC is plugged before each convolutional layer of our generators (). In most of our experiments, we use a ResNet-like architecture (He et al. (2016)) for with 3 blocks (Gulrajani et al. (2017)) and a final convolutional layer (i.e. convolutional layers, corresponding to 7 WC layers). Fig. 1 shows a simplified block scheme. We do not use WC nor we introduce any novelty in the discriminators (more details in Sec. 5).
Whitening, Coloring, ReLU,Whitening, Coloring, ReLU, . (b) A cWC layer.
Comparing Eq. 1 with Eq. 4, the only difference lies in the class-specific and parameters, which are used to project each standardized feature into a class-specific univariate Gaussian distribution. We propose to replace Eq. 4 with our conditional Whitening and Coloring transform (cWC): and , where:
The first term in Eq. 5 is based on class-specific learnable parameters and . Analogously to (Dumoulin et al. (2016b)), in the forward-pass (both at training and at inference time), conditional information , input to , is used to select the correct
pair to apply. Similarly, during training, gradient information is backpropagated throughonly the corresponding pair used in the forward pass. is the same procedure used in the unconditional case, where all the elements in (independently of their class label) are used to compute and .
The second term in Eq. 5 is based on class-agnostic learnable parameters and , similarly to the unconditional coloring transform in Eq. 2. This second term is used because the class-specific parameters are trained with less data then the corresponding class-agnostic weights. In fact, in a generative task with classes and a training protocol with batches of elements, on average only instances in are associated with label . Thus, on average, at each training iteration, the parameters are updated using an effective batch of elements versus a batch of elements used for all the other class-independent parameters. The class-agnostic term is then used to enforce sharing of the parameters of the overall coloring transform. In Sec. 5.2.1 we empirically show the importance of this second term. Fig. 1 shows the two separate network branches corresponding to the two terms in Eq. 5.
When the number of classes is small and the number of real training images is relatively large, learning a separate for each is not a problem. However, with a large and/or few training data, this may lead to an over-complex generator network which can be hard to train. Indeed, the dimensionality of each is and learning parameters may be difficult with a large .
In order to solve this issue, when is large, we compute using a weighted sum over the elements of a class-independent dictionary . Each row of contains a flattened version of a matrix. More formally, is an matrix, with . The -th row in is given by the concatenation of -dimensional filters: , being a coloring filter for the -th output dimension. is shared over all the classes. Given as input, is computed using:
where is a one-hot representation of class (i.e., a vector of zero elements except one in the -th component) and it is used to select the -th row () of the association matrix . During training, given , only the weights in are updated (i.e., all the other rows , with are not used and do not change). However, all the elements in are updated independently of .
We split our experiments in two main scenarios: unconditional and conditional image generation. In each scenario we first compare our approach with the state of the art and then we analyze the impact of each element of our proposal. We use 5 datasets: CIFAR-10, CIFAR-100 (Torralba et al. (2008)), STL-10 (Coates et al. (2011)
), ImageNet (Deng et al. (2009)) and Tiny ImageNet111https://tiny-imagenet.herokuapp.com/. The latter is composed of images taken from 200 ImageNet categories (500 images per class). In our evaluation we use the two most common metrics for image generation tasks: Inception Score (IS) (Salimans et al. (2016)) (the higher the better) and the Fréchet Inception Distance (FID) (Heusel et al. (2017)) (the lower the better). In the Appendix we also show: (1) a comparison based on human judgments as an additional evaluation criterion, (2) qualitative results for all the datasets used in this section, (3) additional experiments on other smaller datasets.
In our experiments we use different generator and discriminator architectures in combination with different discriminator training protocols and we plug our WC/cWC in the generator of each adopted framework. The goal is to show that our whitening-coloring transforms can consistently improve the quality of the results with respect to different basic GAN networks, training protocols and image generation tasks. Note that, if not otherwise explicitly mentioned, we usually adopt the same hyper-parameter setting of the framewoks we compare with. Altough better results can be achieved using, for instance, an ad-hoc learning rate policy (see Sec. 5.2 and Tab. 2 (right)), we want to emphasize that our whitening-coloring transforms can be easily plugged into existing systems.
In all our experiments, , and , as well as the simplified baselines presented in the ablation studies, are applied before each convolutional layer of our generators.
In the unconditional scenario we use two different training strategies for the discriminator: (1) WGAN with Gradient Penalty (Gulrajani et al. (2017)) (GP) and (2) GAN with Spectral Normalization (Miyato et al. (2018)) (SN). Remind that our WC is used only in the generator (Sec. 3). When WC is used in combination with GP, we call it WC GP, while, in combination with SN, we refer to it as WC SN. Moreover, we use either ResNet (He et al. (2016); Gulrajani et al. (2017)) or DCGAN (Radford et al. (2015)) as the basic generator architecture. In order to make the comparison with the corresponding GP and SN frameworks as fair as possible and to show that the quantitative result improvements are due only to our batch normalization approach, we strictly follow the architectural and training protocol details reported in (Gulrajani et al. (2017); Miyato et al. (2018)). For instance, in WC SN + DCGAN we use the same DCGAN-based generator and discriminator used in SN + DCGAN (Miyato et al. (2018)), with the same number of feature-map dimension () in each layer in both the generator and the discriminator, the same learning rate policy (which was optimized for SN + DCGAN but not for our framework), etc. We provide implementation details in Sec. C.1.
In Tab. 1 we show that WC improves both the IS and the FID values of all the tested frameworks in both the CIFAR-10 and the STL-10 dataset. Specifically, independently of the basic architecture (either ResNet or DCGAN), our WC transform improves the corresponding SN results. Similarly, our WC GP results are better than the corresponding GP values.
|Method||IS||FID 5k||FID 10k||IS||FID 5k||FID 10K|
|GP + ResNet (Gulrajani et al. (2017))||7.86 .07||-||-||-||-||-|
|WC GP + ResNet (ours)||8.20 .08||-||20.4||-||-||-|
|SN + DCGAN (Miyato et al. (2018))||7.58 .12||25.5||-||8.79 .14||43.2||-|
|WC SN + DCGAN (ours)||7.84 .10||25.5||23.0||9.45 0.18||40.1||37.9|
|SN + ResNet (Miyato et al. (2018))||8.22 .05||21.7||-||9.10 .04||40.1||-|
|WC SN + ResNet (ours)||8.66 .11||20.2||17.2||9.93 .13||38.7||36.4|
In Tab. 2 (left) we compare WC SN + ResNet with other methods on CIFAR-10 and we show that we reach results almost on par with the state of the art on this dataset (Karras et al. (2018)). Note that the average IS values over 10 runs reported in Karras et al. (2018) is exactly the same as our average IS value: 8.56 (the results reported in the table refer to the best values for each framework). Moreover, Karras et al. (2018) use a much higher capacity generator (about the number of parameters of our WC SN + ResNet) and discriminator (about parameters than ours).
Finally, in Fig. 2 we plot the IS and the FID values computed at different mini-batch training iterations. These plots empirically show that the proposed batch normalization approach significantly speeds-up the training process. For instance, the IS value of WC SN + ResNet after 20k iterations is already higher than the IS value of SN + ResNet after 50k iterations.
We use CIFAR-10 to analyze the influence of each WC element and an SN + ResNet framework (Miyato et al. (2018)), omitting ’SN’ and ’ResNet’ from the names for simplicity.
Our initial baseline is W-only, in which we use Eq. 3 to project the features in a spherical distribution without coloring. In WC-diag, after whitening (Eq. 3), we use Eq. 2 with diagonal matrices. This corresponds to learning scalar - parameters for a dimension-wise scaling-shifting transform as in BN (Eq. 1). On the other hand, in std-C we standardize the features as in BN (Eq. 1) but replacing the - parameters with our coloring transform (Eq. 2) and in C-only, Eq. 2 is applied directly to the original, non-normalized batch instances. Finally, WC is our full-pipeline.
Tab. 4 shows that feature whitening without coloring performs poorly. After the feature normalization step, it is crucial for the network to re-project the features in a new (learned) distribution with a sufficient representation capacity. This is done in standard BN using the - parameters. Analogously, in our WC, this role is played by the coloring step. In fact, with a diagonal-matrix based coloring, WC-diag improves the results with respect to W-only. Moreover, a drastic improvement is obtained by the full-pipeline (WC) with respect to WC-diag, showing the importance of learning a full covariance matrix in the coloring step. On the other hand, the comparison between WC with both C-only and std-C shows that feature normalization before coloring is also important, and that the full advantage of coloring is obtained when used in combination with whitening.
In the same table we report the results obtained by replacing our Cholesky decomposition with the ZCA-based whitening (), which is used in (Huang et al. (2018)), keeping all the rest of our method unchanged (coloring included). The comparison with WC shows that our proposed procedure (detailed in Sec. A) achieves better results than . The results reported in Tab. 4 correspond to the best values achieved by during training and in Sec. E we show that can be highly unstable. Moreover, is, on average, much slower than WC. Specifically, the whitening only part, when computed using the procedure proposed in Sec. A, is more than 11 times faster than the ZCA-based whitening, while, considering the whole pipeline (most of which is shared between WC and ), WC is more than 3.5 times faster than . However, note that in the Decorrelated Batch Normalization (DBN) (Huang et al. (2018)), no coloring is used and the ZCA-based whitening is applied to only the first layer of a classification network, trained in a standard supervised fashion. Hence, has to be considered as our proposed WC in which whitening is performed using the procedure proposed in (Huang et al. (2018)).
In the cGAN scenario, conditional information is represented in the generator as explained in Sec. 4 and using either cWC or (Sec. 4.1). The purpose of this section is to show the advantage of using our conditional coloring with respect to other cGAN methods for representing categorical information input to a generator.
In Tab. 5 we use different basic discriminator and training-protocol frameworks and we report results with and without our conditional batch normalization approach in the corresponding generator. Similarly to the experiments in Sec. 5.1, in our cWC and based GANs we follow the implementation details of the reference frameworks and use their hyper-parameters (e.g., the learning rate policy, etc.) which are not tuned for our method. The only difference with respect to the approaches we compare with (apart from our conditional batch normalization) is a different ratio between the number of filters in the discriminator and those used in the generator in the SN + Proj. Discr. experiments (but keeping constant the total number of filters, see Sec. C.2 for more details). Tab. 5 shows that both cWC and improve the IS and the FID values with respect to the basic frameworks in both the CIFAR-10 and the CIFAR-100 dataset. On CIFAR-10, cWC outperforms , probably because, with a small number of classes (), learning class-specific coloring filters () is relatively easy and more informative than using a class-independent dictionary. However, with , as in the case of the Tiny ImageNet dataset, obtains significantly better results than cWC (see Tab. 4). In Tab. 4, cBN corresponds to our re-implementation of (Miyato and Koyama (2018)), with a ResNet generator + cBN (Dumoulin et al. (2016b)) and an SN + ResNet + Projection Discriminator (Miyato and Koyama (2018)). Note that both cWC and significantly improve both the IS and the FID values with respect to cBN on this large and challenging dataset.
In Tab. 2 (right) we compare cWC and with other methods on CIFAR-10, where cWC achieves the best so far published conditioned results on this dataset. Note that cWC SN + Proj. Discr. in Tab. 2 (right) and in Tab. 5 indicate exactly the same framework, the only difference being the learning rate policy (see Sec. C.2). In fact, to emphasize the comparison, in Tab. 5 we adopted the same learning rate policy used by Miyato and Koyama (2018), which has been customized to our approach only in Tab. 2 (right).
Finally, similarly to the experiment shown in Fig. 2, we show in Tab. 6 that leads to a much faster training with respect to cBN. Specifically, we use the ImageNet dataset (Deng et al. (2009)), which corresponds to a very large-scale generation task and we compare with a similar experiment performed by Miyato and Koyama (2018). The first row of Tab. 6 shows the original results reported in (Miyato and Koyama (2018)) for different training iterations. In the second row we show our results. Note that, in order to fit the generator network used by Miyato and Koyama (2018) on a single GPU, we had to reduce its capacity, decreasing the number of convolutional filters. Hence, our generator, including our additional coloring filters, has 6M parameters while the generator of (Miyato and Koyama (2018)) has 45M parameters. The discriminator is the same for both methods: 39M parameters. These results show that, using a generator with only of the parameters of the generator used in (Miyato and Koyama (2018)), can learn much faster and reach higher IS results on a big dataset like ImageNet. The importance of this experiment lies also in the fact that it emphasizes that the advantage of our method does not depend on the increase of the parameters due to the coloring layers.
|Method||IS||FID 5k||FID 10k||IS||FID 5k||FID 10K|
|GP + ACGAN (Gulrajani et al. (2017))||8.42 .10||-||-||-||-||-|
|cWC GP + ACGAN (ours)||8.80 0.13||-||17.2||-||-||-|
|GP + ACGAN (ours)||8.69 0.13||-||16.8||-||-||-|
|SN + Proj. Discr. (Miyato and Koyama (2018))||8.62||17.5||-||9.04||23.2||-|
|cWC SN + Proj. Discr. (ours)||8.97 .11||16.0||13.5||9.52 .07||20.5||17.2|
|SN + Proj. Discr. (ours)||8.85 .10||16.5||13.5||9.80 .17||20.6||17.4|
|SN + Proj. Discr. (Miyato and Koyama (2018))||17.5||21||23.5||29.7 0.61|
|SN + Proj. Discr. (ours)||21.39 0.37||26.11 0.40||29.20 0.87||34.36 0.67|
We analyze all the main aspects of the conditional coloring introduced in Sec. 4 using an SN + ResNet + Proj. Discr. framework for all the baselines presented here. The whitening part (when used) is obtained using Eq. 3. We call cWC-cls-only a version of with only the class-specific term in Eq. 5. This is implemented simply removing the class-agnostic branch in Fig. 1 and learning a specific ( pair per class, without class-filter soft assignment. -cls-only is the same as cWC-cls-only but using the soft assignment in Eq. 6. Similarly, we call cWC-diag the baseline in which each in Eq. 5 is replaced with a diagonal matrix. This corresponds to learning a set of independent scalar parameters for each class () to represent class-specific information similarly to cBN (see Eq. 4). In cWC-diag we also use the class-agnostic branch. Note that a class-agnostic branch without any kind of class-specific parameters cannot represent at all class information in the generator. As a reference, in the first row of Tab. 7 we also report the results of exactly the same framework but with our cWC replaced with cBN (Dumoulin et al. (2016b)). Moreover, we call c-std-C a version of our method in which feature normalization is performed using standardization instead of whitening and then the two coloring branches described in Sec. 4 (Eq. 5) are applied to the standardized features. Finally, c-std-C corresponds to the soft-assignment version of our conditional coloring (Sec. 4.1) used together with feature standardization.
We first analyze the whitening-based baselines (second part of Tab. 7). The results reported in Tab. 7 show that both versions based on a single coloring branch (cls-only) are significantly worse than cWC-diag which includes the class-agnostic branch. On the other hand, while on CIFAR-10 cWC-diag, cWC and reach comparable results, on CIFAR-100 cWC-diag drastically underperforms the other two versions which are based on full-coloring matrices. This shows the representation capacity of our conditional coloring, which is emphasized in CIFAR-100 where the number of classes () is one order of magnitude larger than in CIFAR-10: With a large , the scalar “” parameters are not sufficient to represent class-specific information. Also the comparison between cWC and depends on , and the advantage of using instead of cWC is more evident on the Tiny ImageNet experiment (see Tab.4). Finally, comparing the standardization-based versions c-std-C and c-std-C with the corresponding whitening-based ones (cWC and ), the importance of full-feature decorrelation before coloring becomes clear. Note that c-std-C and c-std-C have exactly the same number of parameters than cWC and , respectively, which confirms that the advantage of our method does not depend only on the increase of the filter number (see also the ImageNet experiment in Tab. 6).
|Method||IS||FID 10k||IS||FID 10K|
|cBN||7.68 .13||22.1||8.08 .08||26.5|
|c-std-C||7.92 .11||24.37||7.67 .08||46.46|
|c-std-C||7.93 .11||26.75||8.23 .10||27.59|
|cWC-cls-only||8.10 .09||28.0||7.10 .06||75.4|
|-cls-only||8.20 .09||23.5||7.88 .12||33.4|
|cWC-diag||8.90 .09||14.2||8.82 .12||20.8|
|cWC||8.97 .11||13.5||9.52 .07||17.2|
|8.85 .10||13.5||9.80 .17||17.4|
In this section we plug our WC into the classification networks used in (Huang et al. (2018)) and we compare with their DBN (Sec. 2) in a discriminative scenario. DBN basically consists of a ZCA-based feature whitening followed by a scaling-and-shifting transform (Sec. 2). Huang et al. (2018) plug their DBN in the first layer of different ResNet architectures (He et al. (2016)) with varying depth. DBN is based also on grouping, which consists in computing smaller (and, hence, more stable) covariance matrices by grouping together different feature dimensions. For a fair comparison, we also plug our WC in only the first layer of each network but we do not use grouping.
Tab. 8 shows the results averaged over 5 runs. Both WC and DBN achieve lower errors than BN in all the configurations. DBN is slightly better than WC, however the absolute error difference is quite marginal. Moreover, note that the maximum error difference over all the tested configurations is lower than . This empirical analysis, compared with the significant boost we obtain using WC and cWC in Sec. 5, shows that full-feature whitening and coloring is more helpful in a GAN scenario than in a discriminative setting. As mentioned in Sec. 1-2, we believe that this is due to the higher instability of GAN training. In fact, feature normalization makes the loss landscape smoother (Santurkar et al. (2018); Kohler et al. (2018)), making the gradient-descent weight updates closer to Newton updates (Huang et al. (2018); LeCun et al. (2012); S. Wiesler (2011)). Since full-feature whitening performs a more principled normalization, this helps adversarial training in which instability issues are more critical than in a discriminative scenario (Odena et al. (2018)).
In this paper we proposed a whitening and coloring transform which extends both BN and cBN for GAN/cGAN training, respectively. In the first case, we generalize BN introducing full-feature whitening and a multivariate feature re-projection. In the second case, we exploit our learnable feature re-projection after normalization to represent significant categorical conditional information in the generator. Our empirical results show that the proposed approach can speed-up the training process and improve the image generation quality of different basic frameworks on different datasets in both the conditional and the unconditional scenario.
We thank the NVIDIA Corporation for the donation of the GPUs used in this project. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 788793-BACKUP).
Whitening and coloring transforms for multivariate gaussian random variables. Project Rhea. Cited by: §1, §1, §3.
Conditional image synthesis with auxiliary classifier GANs. In ICML, pp. 2642–2651. Cited by: §1, Table 2.
80 million tiny images: A large data set for nonparametric object and scene recognition. IEEE Trans. Pattern Anal. Mach. Intell. 30 (11), pp. 1958–1970. Cited by: §5.
Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv:1708.07747. Cited by: Appendix F.
In order to compute the whitening matrix in Eq. 3, we first need to compute the batch-dependent covariance matrix . Since this computation may be unstable, we use the Shrinkage estimator (Schäfer and Strimmer (2005)) which is based on blending the empirical covariance matrix with a regularizing matrix (in our case we use , the identity matrix):
Once is computed, we use the Cholesky decomposition (Dereniowski and Marek (2004)) in order to compute in Eq. 3 such that . Note that the Cholesky decomposition is unique given and that it is differentiable, thus we can back-propagate the gradients through our whitening transform (more details in Sec. B
). Many modern platforms for deep-network developing such as TensorFlowinclude tools for computing the Cholesky decomposition. Moreover, its computational cost is smaller than other alternatives such as the ZCA whitening (Kessy et al. (2017)) used, for instance, in (Huang et al. (2018)). We empirically showed in Sec. 5.1.1 that, when used to train a GAN generator, our whitening transform is much faster and achieves better results than the ZCA whitening.
In more details, is computed using a lower triangular matrix and the following steps:
We start with , which is equivalent to .
We use the Cholesky decomposition in order to compute and from .
We invert and we eventually get which is used in Eq. 3.
At inference time, once the network has been trained, and in Eq. 3 are replaced with the expected and , computed over the whole training set and approximated using the running average as in (Ioffe and Szegedy (2015)). In more detail, at each mini-batch training iteration , we compute an updated version of and by means of:
When training is over, we use and the steps described above to compute (specifically, we replace with in Eq. 7). Note that this operation needs to be done only once, after that and are fixed and can be used for inference.
The computational cost of is given by the sum of computing () plus the Cholesky decomposition (), the inversion of () and the application of Eq. 3 to the samples in (). Note that optimized matrix multiplication can largely speed-up most of the involved operations. On the other hand, is implemented as a convolutional operation which can also be largely sped-up using common GPUs. Overall, WC is . Using a ResNet architecture with 7 WC layers, in our CIFAR-10 experiments ( images), we need 0.82 seconds per iteration on average on a single NVIDIA Titan X. This includes both the forward and the backward pass for 1 generator and 5 discriminator updates (Gulrajani et al. (2017)). In comparison, the standard BN, with the same architecture, takes on average 0.62 seconds. The ratio WC/BN is 1.32, which is a minor relative overhead. At inference time, since and are pre-computed, WC is reduced to the application of Eq. 3 and the convolutional operations implementing coloring.
Concerning the conditional case, the computational cost of both versions (cWC and ) is basically the same as WC. In , the only additional overhead of the class-specific branch is given by the computation of , which is . We again point out that optimized matrix multiplication can largely speed-up most of the involved operations.
In our implementation, the gradient backpropagation through our WC/cWC layers is obtained by relying on the TensorFlow automatic differentiation. However, for completeness and to make our method fully-reproducible, in this section we provide a reverse mode automatic differentiation (AD) (Giles (2008); Ioffe and Szegedy (2015); Walter (2012)) of the back-propagation steps. Note that, since (conditional and unconditional) is implemented using convolution operations, back-propagation for is straightforward.
In the rest of this section we adopt the notation used by Giles (2008) that we shortly present below. Given a scalar function of a matrix , indicates the reverse mode AD sensitivities with respect to , i.e.: .
Given a batch , we subtract the mean from each sample . Reverse mode AD mean-subtraction is trivial and described in (Ioffe and Szegedy (2015)). We then stack all the mean-subtracted samples in a matrix (where we drop the subscript for simplicity). Using this notation, the empirical covariance matrix in Eq. 7 is given by: .
In the Cholesky decomposition we have: , . We can rewrite Eq. 3 as a matrix multiplication and obtain , where i-th column of is .
In reverse mode AD, is given and we need to compute . Using the formulas in (Giles (2008)) we obtain:
Now, using the formulas in (Walter (2012)):
and is Hadamard product.
Referring again to (Giles (2008)), and combining the reverse sensitives with respect to from the 2 streams we get:
In our experiments we use 2 basic network architectures: ResNet (Gulrajani et al. (2017)) and DCGAN (Miyato et al. (2018)). We strictly follow the training protocol details reported in (Gulrajani et al. (2017); Miyato et al. (2018)) and provided in the two subsections below. Moreover, we adopt the same architectural details suggested in (Gulrajani et al. (2017); Miyato et al. (2018)) (e.g., the number of blocks, the number of convolutional filters, etc.), except the use of our WC/cWC layers in our generators. We use the same hyper-parameter setting of the works we compare with, a choice which likely favours the baselines used for comparison. The only difference with respect to the architectural setup in (Gulrajani et al. (2017); Miyato et al. (2018)) is emphasized in Sec. C.2 and concerns a different ratio between the number of convolutional filters in the generator versus the number of discriminator’s convolutional filters in the SN + Proj. Discr experiments.
In all our -based experiments we use Adam with learning rate , first momentum and second momentum . The learning rate is linearly decreased till it reaches 0. In all the -based experiments we use the WGAN loss + Gradient Penalty (Gulrajani et al. (2017)), while in the -based experiment we use the hinge loss (Miyato et al. (2018)). In all the experiments we train the networks for 100k iterations following (Gulrajani et al. (2017)). In the CIFAR-10 experiments, SN + ResNet was trained for 50k iterations (Miyato et al. (2018)). In the STL-10 experiments, we train the networks for 100k iterations following (Miyato et al. (2018)). Each iteration is composed of 1 generator update with batch size and 5 discriminator updates with batch size (i.e., generated and real images), same numbers are used in (Gulrajani et al. (2017); Miyato et al. (2018)). When training the DCGAN architectures we use two times more iterations. However, in DCGAN each iteration is composed of 1 generator update with batch size and 1 discriminator update with batch size (Miyato et al. (2018)). In all the experiments we use the original image resolution, except for STL-10 where we downsample all images to following (Miyato and Koyama (2018)).
We provide below the architectural details of our networks. We denote with a block consisting of the following layer sequence: (1) either a WC or a cWC layer (depending on the unconditional/conditional scenario, respectively), (2) a ReLU and (3) a convolutional layer with output filters. The output of this block has the same () spatial resolution of the input. is similar to , but the output feature map is downsampled (): the convolutional layer uses . is an upconvolution () block with fractional . We use a similar notation for the ResNet blocks: is a ResNet block (see Fig.1) which preserves spatial resolution. downsamples the feature maps and upsamples the feature maps. Also we denote with: (1) a dense layer with outputs, (2) a global-sum-pooling layer, (3) a flattening layer, (4) a layer that permutes the dimensions of the input feature map to the target spatial resolution . Given the above definitions, we show in Tab. 9 the architectures used for the unconditional image generation experiments.
We again emphasize that the methods we compare with in Tab. 1 have the same network structure except the use of WC layers in the generator.
|WC GP + ResNet|
|WC SN + DCGAN|
|WC SN + ResNet|
|WC SN + DCGAN|
|WC SN + ResNet|
|CIFAR-10 and CIFAR-100|
|cWC GP + ACGAN|
|cWC SN + Proj. Discr.|
|cWC SN + Proj. Discr.|
|cWC SN + Proj. Discr.|
In the conditional image generation experiments, most of the hyper-parameter values are the same as in Sec. C.1. In the CIFAR-10 experiments with , we train our networks for 100k iterations following (Gulrajani et al. (2017)). In the CIFAR-10 and CIFAR-100 experiments with and Projection Discriminator we train our networks for 50k iterations following (Miyato and Koyama (2018)). In the Tiny ImageNet experiments, we train our networks for 100k iterations. In the ImageNet experiment, we train our networks for 450k iterations with a starting learning rate equal to . All the other hyper-parameters are the same as in Sec. C.1. Tab. 10 shows the architectures used in the conditional image generation experiments.
Both cWC SN + Proj. Discr. and SN + Proj. Discr. used in CIFAR-10 and CIFAR-100 (see Tab. 5) have an opposite generator/discriminator filter-number ratio with respect to the ratio used in the original SN + Proj. Discr. by Miyato and Koyama (2018) ( vs. ). In fact, we decrease the number of convolutional filters in each generator block from 256, used in (Miyato and Koyama (2018)), to 128. Simultaneously, we increase the number of convolutional filters in the discriminator from 128 (Miyato and Koyama (2018)) to 256. This is the only difference between our cWC-based frameworks and (Miyato and Koyama (2018)) in the experiments of Tab. 5. The reason behind this architectural choice is that we hypothesize that our cWC generators are more powerful than the corresponding discriminators. Using the same generator/discriminator filter ratio as in (Miyato and Koyama (2018)), we perform on par with SN + Proj. Discr. (Miyato and Koyama (2018)). In Tab. 7, cBN refers to SN + Proj. Discr. with the same aforementioned filter ratio and used in the cWC baselines.
Finally, in Tab. 2 (right) and in Tab. 5, cWC SN + Proj. Discr. indicates exactly the same framework, the only difference being a different learning rate policy. The results of cWC SN + Proj. Discr. reported in Tab. 5 have been obtained using the same learning rate policy used by Miyato and Koyama (2018). However, our model already achieved its peak performance after 30k iterations. Hence, we decreased the learning rate by a factor of 10 at iteration 30k and then we trained cWC SN + Proj. Discr. for additional 1k iterations. This brought to a performance boost, as reported in Tab. 2 (right).
In our experiments we use 2 evaluation metrics: Inception Score (IS) (Salimans et al. (2016)) and the Fréchet Inception Distance (FID) (Heusel et al. (2017)). These metrics are the two most widely adopted evaluation criteria for generative methods. While most of generative works report quantitative results using only one of these two criteria, we chose to report them both to show the agreement of these two metrics: in most of the presented comparisons and ablation studies, when IS is larger, FID is lower.
IS was originally introduced by Salimans et al. (2016), who suggest to measure IS using 5k generated images, repeat this generation 10 times and then average the results. We follow exactly the same protocol.
FID is a more recent metric and the evaluation protocol is less consolidated. Miyato and Koyama (2018) suggest to compute FID using 10k images from the test set together with 5k generated images. Heusel et al. (2017) suggest to use a sample sizes of at least 10k real and 10k generated images. In all our tables, FID 5k refers to the former protocol (Miyato and Koyama (2018)) and FID 10k refers to the second protocol (Heusel et al. (2017)).
In order to further emphasize the improvement we have when using instead of cBN, we show in Tab. 11 a user study performed with 21 persons. Specifically, the dataset used is Tiny ImageNet and the cBN baseline is the same used for the experiments reported in Tab. 4. We show to each user 50 paired image-sets (using 50 randomly selected classes), generated using either or cBN. Each image-set consists of 16 stacked images of the same class, used to increase the robustness of the user’s choice. The user selects the image-set which he/she believes is more realistic. These results clearly show the superiority of against cBN.
|User-averaged class-based||Cronbach’s alpha|
|preference for||inter-user agreement|
|cBN vs||67.34 %||0.87 (Good)|
In this section we compare the stability properties of our Cholesky decomposition with the ZCA-based whitening proposed in (Huang et al. (2018)). Specifically, we use which is the variant of our method introduced in Sec. 5.1.1, where we replace our Cholesky decomposition with the ZCA-based whitening. In Fig. 4 we plot the IS/training-iteration curves corresponding to both and . This graph shows that is highly unstable, since training can frequently degenerate. When this happens, collapses to a model that always produces a constant, uniform grey image.
The reason for this behaviour is likely related to the way in which the gradients are backpropagated through the ZCA-based whitening step. Specifically, following (Huang et al. (2018)
), the gradient of the loss with respect to the covariance matrix depends from the inverse of the difference of the covariance-matrix singular values (see (Huang et al. (2018)), Appendix A.2). As a consequence, if some of the singular values are identical or very close to each other, this computation is ill-conditioned. Conversely, our whitening procedure based on the Cholesky decomposition (Sec. A) has never showed these drastic instability phenomena (Fig. 4).
In this section we show additional experiments using two small-size datasets: the MNIST (LeCun et al. (1998)) and the Fashion-MNIST dataset (Xiao et al. (2017)) (see Fig. 9 for some generated image examples). In both cases we use the SN + ResNet architecture described in Tab. 12. On MNIST we train for 5k iterations with a constant learning rate, while on Fashion-MNIST we train for 20k iterations with a linearly decaying learning rate (see Sec. C.1). All the other hyper-parameters are the same as in Sec. C.1.
The results are reported in Tab. 13. Unconditional image generation on MNIST is the only setup in which we have not observed an improvement using our WC. However, on the more challenging Fashion-MNIST dataset, we did observe an improvement, especially in the conditional scenario.
In this section we show some qualitative results obtained using our WC/cWC methods. Specifically, examples of images generated using the CIFAR-10 dataset are shown in Fig. 5. Fig. 6 (a)-6 (b) show examples generated on the STL-10 and the CIFAR-100 dataset, respectively. In Fig. 8-8 we show Tiny ImageNet and ImageNet samples, respectively. Fig. 9 shows MNIST and Fashion-MNIST generated images.
Finally, in Fig. 7 we show some images generated using CIFAR-10 and CIFAR-100 in which the value of the noise vector is kept fixed while varying the conditioning class .