# Introduction In this coursework we present two variants of the GAN architecture - DCGAN and CGAN, applied to the MNIST dataset and evaluate performance metrics across various optimisations techniques. The MNIST dataset contains 60,000 training images and 10,000 testing images of size 28x28, spread across ten classes representing the ten handwritten digits. Generative Adversarial Networks present a system of models which learn to output data similar to training data. A trained GAN takes noise as an input and is able to provide an output with the same dimensions and relevant features as the samples it has been trained with. GANs employ two neural networks - a *discriminator* and a *generator* which contest in a min-max game. The task of the *discriminator* is to distinguish generated images from real images, while the task of the generator is to produce realistic images which are able to fool the discriminator. Training a shallow GAN with no convolutional layers poses problems such as mode collapse and unbalanced G-D losses which lead to low quality image output. \begin{figure} \begin{center} \includegraphics[width=16em]{fig/generic_gan_mode_collapse.pdf} \caption{Vanilla GAN mode collapse} \label{fig:mode_collapse} \end{center} \end{figure} Some of the main challanges faced when training a GAN are: **mode collapse**, **low quality** of images and **mismatch** between generator and discriminator loss. Mode collapse is achieved with our naive *vanilla GAN* (Appendix-\ref{fig:vanilla_gan}) implementation after 200,000 batches. The generated images observed during a mode collapse can be seen in figure \ref{fig:mode_collapse}. The output of the generator only represents few of the labels originally fed. When mode collapse is reached the loss function of the generator stops improving as shown in figure \ref{fig:vanilla_loss}. We observe the discriminator loss tends to zero as the discriminator learns to assume and classify the fake 1s, while the generator is stuck producing 1 and hence not able to improve. An improvement to the vanilla architecture is Deep Convolutional Generative Adversarial Networks (DCGAN). # DCGAN ## DCGAN Architecture description DCGAN exploits convolutional stride to perform downsampling and transposed convolution to perform upsampling. We use batch normalization at the output of each convolutional layer (exception made for the output layer of the generator and the input layer of the discriminator). The activation functions of the intermediate layers are `ReLU` (for generator) and `LeakyReLU` with slope 0.2 (for discriminator). The activation functions used for the output are `tanh` for the generator and `sigmoid` for the discriminator. The convolutional layers' output in the discriminator uses dropout before feeding the next layers. We noticed a significant improvement in performance, and estimated an optimal dropout rate of 0.25. The optimizer used for training is `Adam(learning_rate=0.002, beta=0.5)`. The main architecture used can be observed in figure \ref{fig:dcganarc}. \begin{figure} \begin{center} \includegraphics[width=24em]{fig/DCGAN_arch.pdf} \caption{DCGAN Architecture} \label{fig:dcganarc} \end{center} \end{figure} ## Tests on MNIST We evaluate three different GAN architectures, varying the size of convolutional layers in the generator, while retaining the structure presented in figure \ref{fig:dcganarc}: * Shallow: Conv128-Conv64 * Medium: Conv256-Conv128 * Deep: Conv512-Conv256 \begin{figure} \begin{center} \includegraphics[width=24em]{fig/med_dcgan_ex.png} \includegraphics[width=24em]{fig/med_dcgan.png} \caption{Medium DCGAN} \label{fig:dcmed} \end{center} \end{figure} We observed that the deep architectures result in a more easily achievable equilibria of G-D losses. Our medium depth DCGAN achieves very good performance (figure \ref{fig:dcmed}), balancing both binary cross entropy losses at approximately 0.9 after 5,000 batches, reaching equilibrium quicker and with less oscillation than the Deepest DCGAN tested (figure \ref{fig:dclong}). As DCGAN is trained with no labels, the generator's primary objective is to output images that fool the discriminator, but does not intrinsically separate the classes from each another. Therefore we sometimes observe oddly shaped digits which may temporarily be labeled as real by the discriminator. This issue is solved by training the network for more batches or introducing a deeper architecture, as it can be deducted from a qualitative comparison between figures \ref{fig:dcmed}, \ref{fig:dcshort} and \ref{fig:dclong}. Applying Virtual Batch Normalization our Medium DCGAN does not provide observable changes in G-D losses, but reduces within-batch correlation. Although it is difficult to qualitatively assess the improvements, figure \ref{fig:vbn_dc} shows results of the introduction of this technique. We evaluated the effect of different dropout rates (results in appendix figures \ref{fig:dcdrop1_1}, \ref{fig:dcdrop1_2}, \ref{fig:dcdrop2_1}, \ref{fig:dcdrop2_2}) and concluded that the optimisation of the dropout hyper-parameter is essential for maximising performance. A high dropout rate results in DCGAN producing only artifacts that do not match any specific class due to the generator performing better than the discriminator. Conversely a low dropout rate leads to an initial stabilisation of G-D losses, but ultimately results in instability under the form of oscillation when training for a large number of batches. Trying different parameters for artificial G-D balancing in the training stage did not achieve any significant benefits, exclusively leading to the generation of more artifacts (figure \ref{fig:baldc}). We also attempted to increase the D training steps with respect to G, but no mode collapse was observed even with the shallow model. \begin{figure} \begin{center} \includegraphics[width=12em]{fig/bal4.png} \caption{DCGAN Balancing G-D; D/G=3} \label{fig:baldc} \end{center} \end{figure} # CGAN ## CGAN Architecture description CGAN is a conditional version of a GAN which utilises labeled data. Unlike DCGAN, CGAN is trained with explicitly provided labels which allow CGAN to associate features with specific classes. The baseline CGAN which we evaluate is visible in figure \ref{fig:cganarc}. The baseline CGAN architecture presents a series of blocks, each containing a dense layer, `LeakyReLu` layer (`slope=0.2`) and a Batch Normalisation layer. The baseline discriminator uses Dense layers, followed by `LeakyReLu` (`slope=0.2`) and a Droupout layer. The optimizer used for training is `Adam`(`learning_rate=0.002`, `beta=0.5`). The architecture of the Deep Convolutional CGAN (cDCGAN) analysed is presented in the Appendix. It uses transpose convolutions with a stride of two to perform upscaling followed by convolutional bloks with singular stride. We find that kernel size of 3 by 3 worked well for all four convolutional blocks which include a Batch Normalization and an Activation layer. The architecture assessed in this paper uses multiplying layers to multiply the label embedding with the output `ReLu` blocks, as we found that it was more robust compared to addition of the label embedding via concatenation. The list of the architecture we evaluate in this report: * Shallow CGAN - 1 `Dense-LeakyReLu` blocks * Medium CGAN - 3 `Dense-LeakyReLu` blocks * Deep CGAN - 5 `Dense-LeakyReLu` blocks * Deep Convolutional CGAN (cDCGAN) * One-Sided Label Smoothing (LS) * Various Dropout (DO): 0.1, 0.3 and 0.5 * Virtual Batch Normalisation (VBN) - Normalisation based on one batch(BN) [@improved] \begin{figure} \begin{center} \includegraphics[width=24em]{fig/CGAN_arch.pdf} \caption{CGAN Architecture} \label{fig:cganarc} \end{center} \end{figure} ## Tests on MNIST When comparing the three levels of depth for the baseline architecture it is possible to notice significant differences in G-D losses balancing. In a shallow architecture we notice a high oscillation of the generator loss (figure \ref{fig:cshort}), which is being overpowered by the discriminator. Despite this we don't experience any issues with vanishing gradient, hence no mode collapse is reached. Similarly, with a deep architecture the discriminator still overpowers the generator, and an equilibrium between the two losses is not achieved. The image quality in both cases is not really high: we can see that even after 20,000 batches some pictures appear to be slightly blurry (figure \ref{fig:clong}). The best compromise is reached for `3 Dense-LeakyReLu` blocks as shown in figure \ref{fig:cmed}. It is possible to observe that G-D losses are perfectly balanced, and their value goes below 1. The image quality is better than the two examples reported earlier, proving that this Medium-depth architecture is the best compromise. \begin{figure} \begin{center} \includegraphics[width=24em]{fig/med_cgan_ex.png} \includegraphics[width=24em]{fig/med_cgan.png} \caption{Medium CGAN} \label{fig:cmed} \end{center} \end{figure} Unlike DCGAN, the three levels of dropout rate attempted do not affect the performance significantly, and as we can see in figures \ref{fig:cg_drop1_1} (0.1), \ref{fig:cmed}(0.3) and \ref{fig:cg_drop2_1}(0.5), both image quality and G-D losses are comparable. The biggest improvement in performance is obtained through one-sided label smoothing, shifting the true labels form 1 to 0.9 to reinforce discriminator behaviour. Using 0.1 instead of zero for the fake labels does not improve performance, as the discriminator loses incentive to do better (generator behaviour is reinforced). Performance results for one-sided labels smoothing with `true_labels = 0.9` are shown in figure \ref{fig:smooth}. \begin{figure} \begin{center} \includegraphics[width=24em]{fig/smoothing_ex.png} \caption{One sided label smoothing} \label{fig:smooth} \end{center} \end{figure} Virtual Batch normalization does not affect performance significantly. Applying this technique to both the CGAN architectures used keeps G-D losses mostly unchanged. The biggest change we expect to see is a lower correlation between images in the same batch. This aspect will mostly affect performance when training a classifier with the generated images from CGAN, as we will obtain more diverse images. Training with a larger batch size would show more significant results, but since we set this parameter to 128 the issue of within-batch correlation is limited. Similarly to DCGAN, changing the G-D steps did not lead to good quality results as it can be seen in figure \ref{fig:cbalance}, in which we tried to train with D/G=15 for 10,000 batches, trying to initialize good discriminator weights, to then revert to a D/G=1, aiming to balance the losses of the two networks. Even in the case of a shallow network, in which mode collapse should have been more likely, we observed diversity between the samples produced for the same classes, indicating that mode collapse still did not occur. \begin{figure} \begin{center} \includegraphics[width=8em]{fig/bal1.png} \includegraphics[width=8em]{fig/bal2.png} \includegraphics[width=8em]{fig/bal3.png} \caption{CGAN G-D balancing results} \label{fig:cbalance} \end{center} \end{figure} The best performing architecture was cDCGAN. It is difficult to assess any potential improvement at this stage, since the samples produced between 8,000 and 13,000 batches are almost indistinguishable from the ones of the MNIST dataset (as it can be seen in figure \ref{fig:cdc}, middle). Training cDCGAN for more than 15,000 batches is however not beneficial, as the discriminator will keep improving, leading the generator loss to increase and produce bad samples as shown in the reported example. We find a good balance for 12,000 batches. \begin{figure} \begin{center} \includegraphics[width=8em]{fig/cdc1.png} \includegraphics[width=8em]{fig/cdc2.png} \includegraphics[width=8em]{fig/cdc3.png} \caption{cDCGAN outputs; 1000 batches - 12000 batches - 20000 batches} \label{fig:cdc} \end{center} \end{figure} Oscillation on the generator loss is noticeable in figure \ref{fig:cdcloss} due to the discriminator loss approaching zero. One possible adjustment to tackle this issue was balancing G-D training steps, opting for G/D=3, allowing the generator to gain some advantage over the discriminator. This technique allowed to smooth oscillation while producing images of similar quality. Using G/D=6 dampens oscillation almost completely leading to the vanishing discriminator's gradient issue. Mode collapse occurs in this specific case as shown on figure \ref{fig:cdccollapse}. Checking the embeddings extracted from a pretrained LeNet classifier (figure \ref{fig:clustcollapse}) we observe low diversity between features of each class, that tend to collapse to very small regions. \begin{figure} \begin{center} \includegraphics[width=8em]{fig/cdcloss1.png} \includegraphics[width=8em]{fig/cdcloss2.png} \includegraphics[width=8em]{fig/cdcloss3.png} \caption{cDCGAN G-D loss; Left G/D=1; Middle G/D=3; Right G/D=6} \label{fig:cdcloss} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=8em]{fig/cdc_collapse.png} \includegraphics[width=8em]{fig/cdc_collapse.png} \includegraphics[width=8em]{fig/cdc_collapse.png} \caption{cDCGAN G/D=6 mode collapse} \label{fig:cdccollapse} \end{center} \end{figure} Virtual Batch Normalization on this architecture was not attempted as it significantly increased the training time (about twice more). Introducing one-sided label smoothing produced very similar results (figure \ref{fig:cdcsmooth}), hence a quantitative performance assessment will need to be performed in the next section to state which ones are better(through Inception Scores). # Inception Score Inception score is calculated as introduced by Tim Salimans et. al [@improved]. However as we are evaluating MNIST, we use LeNet-5 [@lenet] as the basis of the inception score. We use the logits extracted from LeNet: $$ \textrm{IS}(x) = \exp(\mathbb{E}_x \left( \textrm{KL} ( p(y\mid x) \| p(y) ) \right) ) $$ We further report the classification accuracy as found with LeNet. For coherence purposes the inception scores were calculated training the LeNet classifier under the same conditions across all experiments (100 epochs with `SGD`, `learning rate=0.001`). \begin{table}[H] \begin{tabular}{llll} & Accuracy & IS & GAN Tr. Time \\ \hline Shallow CGAN & 0.645 & 3.57 & 8:14 \\ Medium CGAN & 0.715 & 3.79 & 10:23 \\ Deep CGAN & 0.739 & 3.85 & 16:27 \\ \textbf{cDCGAN} & \textbf{0.899} & \textbf{7.41} & 1:05:27 \\ Medium CGAN+LS & 0.749 & 3.643 & 10:42 \\ cDCGAN+LS & 0.846 & 6.63 & 1:12:39 \\ CCGAN-G/D=3 & 0.849 & 6.59 & 48:11 \\ CCGAN-G/D=6 & 0.801 & 6.06 & 36:05 \\ Medium CGAN DO=0.1 & 0.761 & 3.836 & 10:36 \\ Medium CGAN DO=0.5 & 0.725 & 3.677 & 10:36 \\ Medium CGAN+VBN & 0.735 & 3.82 & 19:38 \\ Medium CGAN+VBN+LS & 0.763 & 3.91 & 19:43 \\ *MNIST original & 0.9846 & 9.685 & N/A \\ \hline \end{tabular} \end{table} ## Discussion ### Architecture We observe increased accruacy as we increase the depth of the GAN arhitecture at the cost of training time. There appears to be diminishing returns with the deeper networks, and larger improvements are achievable with specific optimisation techniques. cDCGAN achieves improved performance in comparison to the other cases analysed as we expected from the results obtained in the previous section, since the samples produced are almost identical to the ones of the original MNIST dataset. ### One Side Label Smoothing One sided label smoothing involves relaxing our confidence on the labels in our data. Tim Salimans et. al. [@improved] show smoothing of the positive labels reduces the vulnerability of the neural network to adversarial examples. We observe significant improvements to the Inception score and classification accuracy in the case of our baseline (Medium CGAN). This technique however did not improve the performance of cDCGAN any further, suggesting that reinforcing discriminator behaviour does not benefit the system in this case. ### Virtual Batch Normalisation Virtual Batch Normalisation is a further optimisation technique proposed by Tim Salimans et. al. [@improved]. Virtual batch normalisation is a modification to the batch normalisation layer, which performs normalisation based on statistics from a reference batch. We observe that VBN improved the classification accuracy and the Inception score due to the provided reduction in intra-batch correlation. ### Dropout Despite the difficulties in judging differences between G-D losses and image quality, dropout rate seems to have a noticeable effect on accuracy and inception score, with a variation of 3.6% between our best and worst dropout cases. Ultimately, judging from the measurements, it is preferable to use a low dropout rate (0.1 seems to be the one that achieves the best results). ### G-D Balancing on cDCGAN Despite achieving lower losses oscillation, using G/D=3 to incentivize generator training did not improve the performance of cDCGAN as it is observed from the inception score and testing accuracy. We obtain in fact 5% less test accuracy, meaning that using this technique in our architecture produces on average lower quality images when compared to our standard cDCGAN. # Re-training the handwritten digit classifier *In the following section the generated data we use will be exclusively produced by our cDCGAN architecture.* ## Results In this section we analyze the effect of retraining the classification network using a mix of real and generated data, highlighting the benefits of injecting generated samples in the original training set to boost testing accuracy. As observed in figure \ref{fig:mix1} we performed two experiments for performance evaluation: * Keeping the same number of training samples while just changing the ratio of real to generated data (55,000 samples in total). * Keeping the whole training set from MNIST and adding generated samples from cDCGAN. \begin{figure} \begin{center} \includegraphics[width=12em]{fig/mix_zoom.png} \includegraphics[width=12em]{fig/added_generated_data.png} \caption{Mix data, left unchanged samples number, right added samples} \label{fig:mix1} \end{center} \end{figure} Both experiments show that training the classification network with the injection of generated data (between 40% and 90%) causes on average a small increase in accuracy of up to 0.2%. In absence of original data the testing accuracy drops significantly to around 40% for both cases. ## Adapted Training Strategy For this section we will use 550 samples from MNIST (55 samples per class). Training the classifier yields major challenges, since the amount of samples available is relatively small. Training for 100 epochs, similarly to the previous section, is clearly not enough. The MNIST test set accuracy reached in this case is only 62%, while training for 300 epochs we can reach up to 88%. The learning curve in figure \ref{fig:few_real} suggests we cannot achieve much better with this very small amount of data, since the validation accuracy plateaus, while the training accuracy almost reaches 100%. We conduct one experiment, feeding the test set to a LeNet trained exclusively on data generated from our CGAN. It is noticeable that training for the first 20 epochs gives good results before reaching a plateau (figure \ref{fig:fake_only}) when compared to the learning curve obtained when training the network with only the few real samples. This indicates that we can use the generated data to train the first steps of the network (initialize weights) and train only with the real samples for 300 epochs to perform fine tuning. As observed in figure \ref{fig:few_init} the first steps of retraining will show oscillation, since the fine tuning will try and adapt to the newly fed data. The maximum accuracy reached before the validation curve plateaus is 88.6%, indicating that this strategy proved to be somewhat successful at improving testing accuracy. We try to improve the results obtained earlier by retraining LeNet with mixed data: few real samples and plenty of generated samples (160,000) (learning curve show in figure \ref{fig:training_mixed}). The peak accuracy reached is 91%. We then try to remove the generated samples to apply fine tuning, using only the real samples. After 300 more epochs (figure \ref{fig:training_mixed}) the test accuracy is boosted to 92%, making this technique the most successful attempt of improvement while using a limited amount of data from MNIST dataset. \begin{figure} \begin{center} \includegraphics[width=12em]{fig/training_mixed.png} \includegraphics[width=12em]{fig/fine_tuning.png} \caption{Retraining; Mixed initialization left, fine tuning right} \label{fig:training_mixed} \end{center} \end{figure} Examples of misclassification are displayed in figure \ref{fig:retrain_fail}. It is visible from a cross comparison between these results and the precision-recall curve displayed in figure \ref{fig:pr-retrain} that the network we trained performs really well for most of the digits, but the low confidence on digit $8$ lowers the overall performance. \begin{figure} \begin{center} \includegraphics[width=24em]{fig/pr-retrain.png} \caption{Retraining; Precision-Recall Curve} \label{fig:pr-retrain} \end{center} \end{figure} # Bonus Questions ## Relation to PCA Similarly to GANs, PCA can be used to formulate **generative** models of a system. While GANs are trained neural networks, PCA is a definite statistical procedure which perform orthogonal transformations of the data. Both attempt to identify the most important or *variant* features of the data (which we may then use to generate new data), but PCA by itself is only able to extract linearly related features. In a purely linear system, a GAN would be converging to PCA. In a more complicated system, we would indeed to identify relevant kernels in order to extract relevant features with PCA, while a GAN is able to leverage dense and convolutional neural network layers which may be trained to perform relevant transformations. ## Data representation Using the pre-trained classification on real training examples we extract embeddings of 10,000 randomly sampled real test examples and 10,000 randomly sampled synthetic examples using both CGAN and cDCGAN from the different classes. We obtain both a PCA and TSNE representation of our data on two dimensions in figure \ref{fig:features}. It is observable that the network that achieved a good inception score (cDCGAN) produces embeddings that are very similar to the ones obtained from the original MNIST dataset, further strengthening our hypothesis about the performance of this specific model. On the other hand, with non cDCGAN we notice higher correlation between the two represented features for the different classes, meaning that a good data separation was not achieved. This is probably due to the additional blur produced around the images with our simple CGAN model. \begin{figure} \centering \subfloat[][]{\includegraphics[width=.2\textwidth]{fig/pca-mnist.png}}\quad \subfloat[][]{\includegraphics[width=.2\textwidth]{fig/tsne-mnist.png}}\\ \subfloat[][]{\includegraphics[width=.2\textwidth]{fig/pca-cgan.png}}\quad \subfloat[][]{\includegraphics[width=.2\textwidth]{fig/tsne-cgan.png}}\\ \subfloat[][]{\includegraphics[width=.2\textwidth]{fig/pca-cdc.png}}\quad \subfloat[][]{\includegraphics[width=.2\textwidth]{fig/tsne-cdc.png}} \caption{Visualisations: a)MNIST|PCA b)MNIST|TSNE c)CGAN-gen|PCA d)CGAN-gen|TSNE e)cDCGAN-gen|PCA f)cDCGAN-gen|TSNE} \label{fig:features} \end{figure} We have presented the Precision Recall Curve for the MNIST, against that of a Dense CGAN and Convolutional CGAN. While the superior performance of the convolutional GAN is evident, it is interesting to note that the precision curves are similar, specifically the numbers 8 and 9. For both architectures 9 is the worst digit on average, but for higher Recall we find that there is a smaller proportion of extremely poor 8's, which result in lower the digit to the poorest precision. \begin{figure} \centering \subfloat[][]{\includegraphics[width=.22\textwidth]{fig/pr-mnist.png}}\quad \subfloat[][]{\includegraphics[width=.22\textwidth]{fig/pr-cgan.png}}\\ \subfloat[][]{\includegraphics[width=.22\textwidth]{fig/pr-cdc.png}} \caption{Precisional Recall Curves a) MNIST : b) CGAN output c)cDCGAN output} \label{fig:rocpr} \end{figure} ## Factoring in classification loss into GAN Classification accuracy and Inception score can be factored into the GAN to attempt to produce more realistic images. Shane Barrat and Rishi Sharma are able to indirectly optimise the inception score to over 900, and note that directly optimising for maximised Inception score produces adversarial examples [@inception-note]. Nevertheless, a pre-trained static classifier may be added to the GAN model, and its loss incorporated into the loss added too the loss of the GAN. $$ L_{\textrm{total}} = \alpha L_{\textrm{LeNet}} + \beta L_{\textrm{generator}} $$ # References
# Appendix ## DCGAN-Appendix \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/vanilla_gan_arc.pdf} \caption{Vanilla GAN Architecture} \label{fig:vanilla_gan} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/generic_gan_loss.png} \caption{Shallow GAN D-G Loss} \label{fig:vanilla_loss} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/short_dcgan_ex.png} \includegraphics[width=24em]{fig/short_dcgan.png} \caption{Shallow DCGAN} \label{fig:dcshort} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/long_dcgan_ex.png} \includegraphics[width=24em]{fig/long_dcgan.png} \caption{Deep DCGAN} \label{fig:dclong} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/vbn_dc.pdf} \caption{DCGAN Virtual Batch Normalization} \label{fig:vbn_dc} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/dcgan_dropout01_gd.png} \caption{DCGAN Dropout 0.1 G-D Losses} \label{fig:dcdrop1_1} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=14em]{fig/dcgan_dropout01.png} \caption{DCGAN Dropout 0.1 Generated Images} \label{fig:dcdrop1_2} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/dcgan_dropout05_gd.png} \caption{DCGAN Dropout 0.5 G-D Losses} \label{fig:dcdrop2_1} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=14em]{fig/dcgan_dropout05.png} \caption{DCGAN Dropout 0.5 Generated Images} \label{fig:dcdrop2_2} \end{center} \end{figure} ## CGAN-Appendix \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/short_cgan_ex.png} \includegraphics[width=24em]{fig/short_cgan.png} \caption{Shallow CGAN} \label{fig:cshort} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/long_cgan_ex.png} \includegraphics[width=24em]{fig/long_cgan.png} \caption{Deep CGAN} \label{fig:clong} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/cgan_dropout01.png} \caption{CGAN Dropout 0.1 G-D Losses} \label{fig:cg_drop1_1} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=14em]{fig/cgan_dropout01_ex.png} \caption{CGAN Dropout 0.1 Generated Images} \label{fig:cg_drop1_2} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/cgan_dropout05.png} \caption{CGAN Dropout 0.5 G-D Losses} \label{fig:cg_drop2_1} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=14em]{fig/cgan_dropout05_ex.png} \caption{CGAN Dropout 0.5 Generated Images} \label{fig:cg_drop2_2} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=18em]{fig/clustcollapse.png} \caption{cDCGAN G/D=6 Embeddings through LeNet} \label{fig:clustcollapse} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=8em]{fig/cdcsmooth.png} \caption{cDCGAN+LS outputs 12000 batches} \label{fig:cdcsmooth} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/smoothing.png} \caption{CGAN+LS G-D Losses} \label{fig:smoothgd} \end{center} \end{figure} ## cDCGAN Alternative Architecture \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/cdcgen.pdf} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/cdcdesc.pdf} \end{center} \end{figure} ## Retrain-Appendix \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/train_few_real.png} \caption{Training with 550 samples from MNIST only} \label{fig:few_real} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/fake_only.png} \caption{Retraining with generated samples only} \label{fig:fake_only} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=24em]{fig/initialization.png} \caption{Retraining with initialization from generated samples} \label{fig:few_init} \end{center} \end{figure} \begin{figure}[H] \begin{center} \includegraphics[width=12em]{fig/retrain_fail.png} \caption{Retraining failures} \label{fig:retrain_fail} \end{center} \end{figure}