http://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&feed=atom&action=historyConditional Image Synthesis with Auxiliary Classifier GANs - Revision history2024-03-28T12:32:52ZRevision history for this page on the wikiMediaWiki 1.41.0http://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31762&oldid=prevVenktech: /* Model */2017-12-04T14:24:17Z<p><span dir="auto"><span class="autocomment">Model</span></span></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 10:24, 4 December 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l87">Line 87:</td>
<td colspan="2" class="diff-lineno">Line 87:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Model ===</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Model ===</div></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The authors apply their AC-GAN model to the ImageNet [[#References|(Russakovsky et al., 2015)]] dataset. The data have 1000 classes which the authors split into 100 groups of 10. An AC-GAN model is trained on each group of 10 to give results reported for the paper. The authors give some examples of the images generated from this setup. They note that these are selected to show the success of the model, not give a balanced representation of how good it is:</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The authors apply their AC-GAN model to the ImageNet [[#References|(Russakovsky et al., 2015)]] dataset<ins style="font-weight: bold; text-decoration: none;">. The architecture of the generator consists of a series of deconvolution layers that transform the noise and class c into an image</ins>. The data have 1000 classes which the authors split into 100 groups of 10. An AC-GAN model is trained on each group of 10 to give results reported for the paper. The authors give some examples of the images generated from this setup. They note that these are selected to show the success of the model, not give a balanced representation of how good it is:</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>[[File:Figure_1_AC-GAN.JPG|thumb|1000px|center|alt=(Odena et al., 2016) Figure 1: Selected images generated by the AC-GAN model for the ImageNet dataset.|(Odena et al., 2016) Figure 1: Selected images generated by the AC-GAN model for the ImageNet dataset.]]</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>[[File:Figure_1_AC-GAN.JPG|thumb|1000px|center|alt=(Odena et al., 2016) Figure 1: Selected images generated by the AC-GAN model for the ImageNet dataset.|(Odena et al., 2016) Figure 1: Selected images generated by the AC-GAN model for the ImageNet dataset.]]</div></td></tr>
</table>Venktechhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31666&oldid=prevJimit: /* Previous Work */2017-11-28T23:55:10Z<p><span dir="auto"><span class="autocomment">Previous Work</span></span></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 19:55, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l13">Line 13:</td>
<td colspan="2" class="diff-lineno">Line 13:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Of all image synthesis methods (e.g. variational autoencoders, autoregressive models, invertible density estimators), GANs have become one of the most popular and successful due to their flexibility and the ease with which they can be sampled from. A standard GAN framework pits a generative model $G$ against a discriminative adversary $D$. The goal of $G$ is to learn a mapping from a latent space $Z$ to a real space $X$ to produce examples (generally images) indistinguishable from training data. The goal of the $D$ is to iteratively learn to predict when a given input image is from the training set or a synthesized image from $G$. Jointly the models are trained to solve the game-theoretical minimax problem, as defined by [[#References|Goodfellow et al. (2014)]]: $$\underset{G}{\text{min }}\underset{D}{\text{max }}V(G,D)=\mathbb{E}_{X\sim p_{data}(x)}[log(D(X))]+\mathbb{E}_{Z\sim p_{Z}(z)}[log(1-D(G(Z)))]$$</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Of all image synthesis methods (e.g. variational autoencoders, autoregressive models, invertible density estimators), GANs have become one of the most popular and successful due to their flexibility and the ease with which they can be sampled from. A standard GAN framework pits a generative model $G$ against a discriminative adversary $D$. The goal of $G$ is to learn a mapping from a latent space $Z$ to a real space $X$ to produce examples (generally images) indistinguishable from training data. The goal of the $D$ is to iteratively learn to predict when a given input image is from the training set or a synthesized image from $G$. Jointly the models are trained to solve the game-theoretical minimax problem, as defined by [[#References|Goodfellow et al. (2014)]]: $$\underset{G}{\text{min }}\underset{D}{\text{max }}V(G,D)=\mathbb{E}_{X\sim p_{data}(x)}[log(D(X))]+\mathbb{E}_{Z\sim p_{Z}(z)}[log(1-D(G(Z)))]$$</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>While this initial framework has clearly demonstrated great potential, other authors have proposed changes to the method to improve it. Many such papers propose changes to the training process [[#References|(Salimans et al., 2016)]][[#References|(Karras et al., 2017)]], which is notoriously difficult for some problems. Others propose changes to the model itself. [[#References|Mirza & Osindero (2014)]] augment the model by supplying the class of observations to both the generator and discriminator to produce class-conditional samples. According to [[STAT946F17/Conditional Image Generation with PixelCNN Decoders|van den Oord et al. (2016)]], conditioning image generation on classes can greatly improve their quality. Other authors have explored using even richer side information in the generation process with good results [[Learning What and Where to Draw|(Reed et al., 2016)]]. A summary diagram of the difference in the architecture can be seen in the following figure.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">This approach can also be interpreted as two neural networks contesting with each other in a zero-sum game. </ins>While this initial framework has clearly demonstrated great potential, other authors have proposed changes to the method to improve it. Many such papers propose changes to the training process [[#References|(Salimans et al., 2016)]][[#References|(Karras et al., 2017)]], which is notoriously difficult for some problems. Others propose changes to the model itself. [[#References|Mirza & Osindero (2014)]] augment the model by supplying the class of observations to both the generator and discriminator to produce class-conditional samples. According to [[STAT946F17/Conditional Image Generation with PixelCNN Decoders|van den Oord et al. (2016)]], conditioning image generation on classes can greatly improve their quality. Other authors have explored using even richer side information in the generation process with good results [[Learning What and Where to Draw|(Reed et al., 2016)]]. A summary diagram of the difference in the architecture can be seen in the following figure.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>[[FILE: ACGAN.png|center|600px]]</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>[[FILE: ACGAN.png|center|600px]]</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
</table>Jimithttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31637&oldid=prevPkthekka: /* Model */2017-11-28T18:56:04Z<p><span dir="auto"><span class="autocomment">Model</span></span></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:56, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l124">Line 124:</td>
<td colspan="2" class="diff-lineno">Line 124:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>A major attraction of this paper is the impressive quality of samples generated by the model. GANs often generate samples that are locally plausible but globally not realistic (e.g. a generated image of a dog has fur but the overall shape is not distinguishable). As we have seen in this critique, and as acknowledged by the authors, the most impressive samples are not representative of the model's overall performance.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>A major attraction of this paper is the impressive quality of samples generated by the model. GANs often generate samples that are locally plausible but globally not realistic (e.g. a generated image of a dog has fur but the overall shape is not distinguishable). As we have seen in this critique, and as acknowledged by the authors, the most impressive samples are not representative of the model's overall performance.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The model itself is not a very big advancement of the field. It combines two ideas that are both already prevalent in the research without any other justification than that it seems like a natural thing to do. As [https://openreview.net/forum?id=BkDDM04Ke other reviewers] have noted, investigating how much value the proposed model adds by comparing it with other models that only implement one (or neither) of the changes would have made this paper a slightly more interesting read.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The model itself is not a very big advancement of the field. It combines two ideas that are both already prevalent in the research without any other justification than that it seems like a natural thing to do. As [https://openreview.net/forum?id=BkDDM04Ke other reviewers] have noted, investigating how much value the proposed model adds by comparing it with other models that only implement one (or neither) of the changes would have made this paper a slightly more interesting read<ins style="font-weight: bold; text-decoration: none;">. The AC-GAN model can perform semi-supervised learning by ignoring the component of the loss arising from class labels when a label is unavailable for a given training image</ins>.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Another criticism I have about the paper is about how they report their results. To compare with [[#References|Salimans et al. (2016)]] they use Inception Score rather than log-likelihood, which they claim is the standard. Even if their model performed worse by that measure it ought to be included with the caveat they mentioned. The models are evaluated on a different dataset and at a lower spatial resolution than was used for the rest of the paper. By the Inception Score their results are better on average but might not be significantly different given how close they are. Finally, they did not apply the mean MS-SSIM score they developed in this paper to evaluate their model against [[#References|Salimans et al. (2016)]]. This would have been a natural point to make, but instead, they generate four samples from each model as their evidence.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Another criticism I have about the paper is about how they report their results. To compare with [[#References|Salimans et al. (2016)]] they use Inception Score rather than log-likelihood, which they claim is the standard. Even if their model performed worse by that measure it ought to be included with the caveat they mentioned. The models are evaluated on a different dataset and at a lower spatial resolution than was used for the rest of the paper. By the Inception Score their results are better on average but might not be significantly different given how close they are. Finally, they did not apply the mean MS-SSIM score they developed in this paper to evaluate their model against [[#References|Salimans et al. (2016)]]. This would have been a natural point to make, but instead, they generate four samples from each model as their evidence.</div></td></tr>
</table>Pkthekkahttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31631&oldid=prevDmittal at 18:40, 28 November 20172017-11-28T18:40:18Z<p></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:40, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l1">Line 1:</td>
<td colspan="2" class="diff-lineno">Line 1:</td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>'''Abstract:''' "In this paper, we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis. Generative adversarial networks (GANs) are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks contesting with each other in a zero-sum game framework []. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." [[#References | (Odena et al., 2016)]]</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>'''Abstract:''' "In this paper, we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis. Generative adversarial networks (GANs) are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks contesting with each other in a zero-sum game framework [<ins style="font-weight: bold; text-decoration: none;">17</ins>]. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." [[#References | (Odena et al., 2016)]]</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Introduction =</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Introduction =</div></td></tr>
</table>Dmittalhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31629&oldid=prevDmittal: /* References */2017-11-28T18:40:00Z<p><span dir="auto"><span class="autocomment">References</span></span></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:40, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l172">Line 172:</td>
<td colspan="2" class="diff-lineno">Line 172:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># G.L. Grinblat, L.C. Uzal, P.M. Granitto. Class-splitting generative adversarial networks. arXiv preprint [https://arxiv.org/abs/1709.07359 : arXiv:1709.07359].</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># G.L. Grinblat, L.C. Uzal, P.M. Granitto. Class-splitting generative adversarial networks. arXiv preprint [https://arxiv.org/abs/1709.07359 : arXiv:1709.07359].</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># Dosovitskiy, Alexey, and Thomas Brox. "Generating images with perceptual similarity metrics based on deep networks." Advances in Neural Information Processing Systems. 2016.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># Dosovitskiy, Alexey, and Thomas Brox. "Generating images with perceptual similarity metrics based on deep networks." Advances in Neural Information Processing Systems. 2016.</div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"># https://en.wikipedia.org/wiki/Generative_adversarial_network</ins></div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>APA </div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>APA </div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
</table>Dmittalhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31628&oldid=prevDmittal at 18:39, 28 November 20172017-11-28T18:39:21Z<p></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:39, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l1">Line 1:</td>
<td colspan="2" class="diff-lineno">Line 1:</td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>'''Abstract:''' "In this paper, we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." [[#References | (Odena et al., 2016)]]</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>'''Abstract:''' "In this paper, we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis<ins style="font-weight: bold; text-decoration: none;">. Generative adversarial networks (GANs) are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks contesting with each other in a zero-sum game framework []</ins>. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." [[#References | (Odena et al., 2016)]]</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Introduction =</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Introduction =</div></td></tr>
</table>Dmittalhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31623&oldid=prevAmirhk at 18:20, 28 November 20172017-11-28T18:20:50Z<p></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:20, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l60">Line 60:</td>
<td colspan="2" class="diff-lineno">Line 60:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># The metric is not saturated. Scores on their generated data vary across the unit interval. If scores were all very close to zero the metric would not be much use.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># The metric is not saturated. Scores on their generated data vary across the unit interval. If scores were all very close to zero the metric would not be much use.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The second point they raise is that the mean MS-SSIM metric is not intended as a proxy for the entropy of the generator distribution in pixel space. That measure is hard to compute, and in any case is sensitive to trivial changes in the pixels, whereas the true intention of this metric is to measure perceptual similarity.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The second point they raise is that the mean MS-SSIM metric is not intended as a proxy for the entropy of the generator distribution in pixel space. That measure is hard to compute, and in any case is sensitive to trivial changes in the pixels, whereas the true intention of this metric is to measure perceptual similarity<ins style="font-weight: bold; text-decoration: none;">. Another approach is proposed by Dosovitskiy and Brox (2016) where class of loss functions, called deep perceptual similarity metrics (DeePSiM), compute distances between image features extracted (rather than distances in image space) to suggest attained diversity in generated images</ins>.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>== Experimental Results on GAN Properties ==</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>== Experimental Results on GAN Properties ==</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l171">Line 171:</td>
<td colspan="2" class="diff-lineno">Line 171:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Berg, A. C. (2015). Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 211-252.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Berg, A. C. (2015). Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 211-252.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># G.L. Grinblat, L.C. Uzal, P.M. Granitto. Class-splitting generative adversarial networks. arXiv preprint [https://arxiv.org/abs/1709.07359 : arXiv:1709.07359].</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># G.L. Grinblat, L.C. Uzal, P.M. Granitto. Class-splitting generative adversarial networks. arXiv preprint [https://arxiv.org/abs/1709.07359 : arXiv:1709.07359].</div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"># Dosovitskiy, Alexey, and Thomas Brox. "Generating images with perceptual similarity metrics based on deep networks." Advances in Neural Information Processing Systems. 2016.</ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">APA </ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"></ins></div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Online resources ===</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Online resources ===</div></td></tr>
</table>Amirhkhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31622&oldid=prevAravindbk: /* Motivation */2017-11-28T18:18:56Z<p><span dir="auto"><span class="autocomment">Motivation</span></span></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:18, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l4">Line 4:</td>
<td colspan="2" class="diff-lineno">Line 4:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Motivation ===</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Motivation ===</div></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The authors introduce a GAN architecture for generating high resolution images from the ImageNet dataset. They show that this architecture makes it possible to split the generation process into many sub-models. They further suggest that GANs have trouble generating globally coherent images and that this architecture is responsible for the coherence of their samples. They <del style="font-weight: bold; text-decoration: none;">experimentally </del>demonstrate that generating higher resolution images allow the model to encode more class-specific information, making them more visually discriminable than lower resolution images even after they have been resized to the same resolution.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The authors introduce a GAN architecture for generating high resolution images from the ImageNet dataset. They show that this architecture makes it possible to split the generation process into many sub-models. They further suggest that GANs have trouble generating globally coherent images and that this architecture is responsible for the coherence of their samples. They demonstrate <ins style="font-weight: bold; text-decoration: none;">that adding more structure</ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">to the GAN latent space along with a specialized cost function results in higher quality samples and </ins>that generating higher resolution images allow the model to encode more class-specific information, making them more visually discriminable than lower resolution images even after they have been resized to the same resolution.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second half of the paper introduces metrics for assessing visual discriminability and diversity of synthesized images. The discussion of image diversity, in particular, is important due to the tendency for GANs to 'collapse' to only produce one image that best fools the discriminator [[#References|(Goodfellow et al., 2014)]].</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second half of the paper introduces metrics for assessing visual discriminability and diversity of synthesized images. The discussion of image diversity, in particular, is important due to the tendency for GANs to 'collapse' to only produce one image that best fools the discriminator [[#References|(Goodfellow et al., 2014)]].</div></td></tr>
</table>Aravindbkhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31620&oldid=prevAravindbk at 18:08, 28 November 20172017-11-28T18:08:22Z<p></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 14:08, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l1">Line 1:</td>
<td colspan="2" class="diff-lineno">Line 1:</td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>'''Abstract:''' "In this paper we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." [[#References | (Odena et al., 2016)]]</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>'''Abstract:''' "In this paper<ins style="font-weight: bold; text-decoration: none;">, </ins>we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." [[#References | (Odena et al., 2016)]]</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Introduction =</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Introduction =</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Motivation ===</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Motivation ===</div></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The authors introduce a GAN architecture for generating high resolution images from the ImageNet dataset. They show that this architecture makes it possible to split the generation process into many sub-models. They further suggest that GANs have trouble generating globally coherent images<del style="font-weight: bold; text-decoration: none;">, </del>and that this architecture is responsible for the coherence of their samples. They experimentally demonstrate that generating higher resolution images allow the model to encode more class-specific information, making them more visually discriminable than lower resolution images even after they have been resized to the same resolution.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The authors introduce a GAN architecture for generating high resolution images from the ImageNet dataset. They show that this architecture makes it possible to split the generation process into many sub-models. They further suggest that GANs have trouble generating globally coherent images and that this architecture is responsible for the coherence of their samples. They experimentally demonstrate that generating higher resolution images allow the model to encode more class-specific information, making them more visually discriminable than lower resolution images even after they have been resized to the same resolution.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second half of the paper introduces metrics for assessing visual discriminability and diversity of synthesized images. The discussion of image diversity, in particular, is important due to the tendency for GANs to 'collapse' to only produce one image that best fools the discriminator [[#References|(Goodfellow et al., 2014)]].</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second half of the paper introduces metrics for assessing visual discriminability and diversity of synthesized images. The discussion of image diversity, in particular, is important due to the tendency for GANs to 'collapse' to only produce one image that best fools the discriminator [[#References|(Goodfellow et al., 2014)]].</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l73">Line 73:</td>
<td colspan="2" class="diff-lineno">Line 73:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Generated Images are both Diverse and Discriminable ===</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== Generated Images are both Diverse and Discriminable ===</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>A second experiment conducted in the paper aims to investigate how the two metrics they propose interact with each other. They simply calculate the Inception accuracy and mean MS-SSIM score for a batch of generated images from every class in their data and report the correlation between the scores. They find the scores are anti-correlated with $\rho=-0.16$. Because the mean MS-SSIM metric is low for diverse samples, they conclude that accuracy and diversity are actually positively correlated, which is contradictory to the hypothesis that GANs that collapse <del style="font-weight: bold; text-decoration: none;">achieve </del>better sample quality.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>A second experiment conducted in the paper aims to investigate how the two metrics they propose interact with each other. They simply calculate the Inception accuracy and mean MS-SSIM score for a batch of generated images from every class in their data and report the correlation between the scores. They find the scores are anti-correlated with $\rho=-0.16$. Because the mean MS-SSIM metric is low for diverse samples, they conclude that accuracy and diversity are actually positively correlated, which is contradictory to the hypothesis that GANs that collapse <ins style="font-weight: bold; text-decoration: none;">achieves </ins>better sample quality.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>[[File:Figure_10.JPG|thumb|300px|right|alt=(Odena et al., 2016) Figure 10: Mean MS-SSIM scores for 10 ImageNet classes (y-axis) plotted against the number of classes handled by each sub-model (x-axis).|(Odena et al., 2016) Figure 10: Mean MS-SSIM scores for 10 ImageNet classes (y-axis) plotted against the number of classes handled by each sub-model (x-axis).]]</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>[[File:Figure_10.JPG|thumb|300px|right|alt=(Odena et al., 2016) Figure 10: Mean MS-SSIM scores for 10 ImageNet classes (y-axis) plotted against the number of classes handled by each sub-model (x-axis).|(Odena et al., 2016) Figure 10: Mean MS-SSIM scores for 10 ImageNet classes (y-axis) plotted against the number of classes handled by each sub-model (x-axis).]]</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l96">Line 96:</td>
<td colspan="2" class="diff-lineno">Line 96:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The authors compare their model with state-of-the-art results from [[#References| Salimans et al. (2016)]] on the CIFAR-10 dataset at a 32 x 32 resolution. To score the two models they use Inception Score instead of log-likelihood, which they claim is too inaccurate to be reported. Their model achieves a score of $8.25 \pm 0.07$ versus the previous state-of-the-art of $8.09 \pm 0.07$.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The authors compare their model with state-of-the-art results from [[#References| Salimans et al. (2016)]] on the CIFAR-10 dataset at a 32 x 32 resolution. To score the two models they use Inception Score instead of log-likelihood, which they claim is too inaccurate to be reported. Their model achieves a score of $8.25 \pm 0.07$ versus the previous state-of-the-art of $8.09 \pm 0.07$.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>[[#References|Odena et al. (2016)]] argue that the class conditional generator allows $G$ to learn a representation of $Z$ independent of $C$ in section 3, and give evidence of the claim later in section 4.5 by showing that images generated with a fixed latent vector $z$ but different class labels $c$ have similar global structure (e.g. orientation of the subject) but the subjects (bird species) vary according to the label. Interestingly, the background (especially in the top row) also varies with the class label. This can possibly be attributed to the bird species coming from different areas, hence a seagull might be expected to have an ocean background. Clearly here the model benefits from the fact that the authors grouped similar classes together. A more interesting analysis might show the same comparison between different classes, such as birds and forklifts, to see how global structure is encoded across them.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>[[#References|Odena et al. (2016)]] argue that the class conditional generator allows $G$ to learn a representation of $Z$ independent of $C$ in section 3, and give evidence of the claim later in section 4.5 by showing that images generated with a fixed latent vector $z$ but different class labels $c$ have similar global structure (e.g. orientation of the subject) but the subjects (bird species) vary according to the label. Interestingly, the background (especially in the top row) also varies with the class label. This can possibly be attributed to the bird species coming from different areas, hence a seagull might be expected to have an ocean background. Clearly here the model benefits from the fact that the authors grouped similar classes together. A more interesting analysis might show the same comparison between different classes, such as birds and forklifts, to see how <ins style="font-weight: bold; text-decoration: none;">the </ins>global structure is encoded across them.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The authors also include a discussion of whether their model is overfitting the training data. Their first test is to find the nearest neighbour in the training set of a generated image by the L1 measure in pixel space and visually compare the two images. This is a fairly naive approach, since the L1 loss in pixel space is extremely unlikely to identify whether two images are perceptually similar. Here would have been a good place to use the MS-SSIM metric to identify the nearest neighbours, since it is intended to measure perceptual similarity. The images they provide from this analysis are below.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The authors also include a discussion of whether their model is overfitting the training data. Their first test is to find the nearest neighbour in the training set of a generated image by the L1 measure in pixel space and visually compare the two images. This is a fairly naive approach, since the L1 loss in pixel space is extremely unlikely to identify whether two images are perceptually similar. Here would have been a good place to use the MS-SSIM metric to identify the nearest neighbours, since it is intended to measure perceptual similarity. The images they provide from this analysis are below.</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l129">Line 129:</td>
<td colspan="2" class="diff-lineno">Line 129:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>An analysis the authors could have included that was touched upon but not explored in section 4.6 of the paper, and in the [[#Results|Results]] section of this summary, is how the similarity of the classes grouped in each sub-model impact the quality of generated samples. The example I gave above was to compare generated images with the same latent code but very different classes, such as birds and forklifts, to see how the global structure transferred across dissimilar classes.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>An analysis the authors could have included that was touched upon but not explored in section 4.6 of the paper, and in the [[#Results|Results]] section of this summary, is how the similarity of the classes grouped in each sub-model impact the quality of generated samples. The example I gave above was to compare generated images with the same latent code but very different classes, such as birds and forklifts, to see how the global structure transferred across dissimilar classes.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The last point to make about the model section is that the authors make some unsupported claims in their discussions of the model's properties. Specifically, they state that their modification to the standard GAN formulation appears to stabilize training but offer no evidence. Another example is their claim that "AC-GANs learn a representation for $z$ that is independent of class label". They <del style="font-weight: bold; text-decoration: none;">site </del>[[#References|Kingma et al (2014)]] as evidence of this. From my review of that paper, it does not appear that the authors gave evidence for such a claim.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The last point to make about the model section is that the authors make some unsupported claims in their discussions of the model's properties. Specifically, they state that their modification to the standard GAN formulation appears to stabilize training but offer no evidence. Another example is their claim that "AC-GANs learn a representation for $z$ that is independent of <ins style="font-weight: bold; text-decoration: none;">the </ins>class label". They <ins style="font-weight: bold; text-decoration: none;">cite </ins>[[#References|Kingma et al (2014)]] as evidence of this. From my review of that paper, it does not appear that the authors gave evidence for such a claim.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== GAN Quality Metrics ===</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>=== GAN Quality Metrics ===</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The Inception accuracy metric proposed in this paper has the drawback that it is only applicable in a conditional GAN setting<del style="font-weight: bold; text-decoration: none;">, </del>since the standard GAN framework has no ground-truth labels. It is also true that using a pre-trained classifier is only a proxy for determining how much generated images look like the class they are meant to represent<del style="font-weight: bold; text-decoration: none;">, </del>since classifiers are not perfect. Consider the phenomenon of adversarial attacks on classifiers to see this point. However, the advantages the authors list, that the Inception accuracy can be computed on a per-class basis and is easier to interpret than the Inception Score do have merit. The metric does make sense for the task the authors use it for.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The Inception accuracy metric proposed in this paper has the drawback that it is only applicable in a conditional GAN setting since the standard GAN framework has no ground-truth labels. It is also true that using a pre-trained classifier is only a proxy for determining how much generated images look like the class they are meant to represent since classifiers are not perfect. Consider the phenomenon of adversarial attacks on classifiers to see this point. However, the advantages the authors list, that the Inception accuracy can be computed on a per-class basis and is easier to interpret than the Inception Score do have merit. The metric does make sense for the task the authors use it for.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The same can be said for the mean MS-SSIM metric developed in this paper. Visually it appears to be a good indicator of diversity in the GAN's outputs. The authors claim that the mean MS-SSIM is a fast and easy-to-compute metric for perceptual variability and collapsing behaviour in a GAN. It is unclear how fast the metric can be computed since for each class the MS-SSIM has to be computed 100*99 times, once for each pair of images. The authors do not discuss how quickly it can actually be done.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The same can be said for the mean MS-SSIM metric developed in this paper. Visually it appears to be a good indicator of diversity in the GAN's outputs. The authors claim that the mean MS-SSIM is a fast and easy-to-compute metric for perceptual variability and collapsing behaviour in a GAN. It is unclear how fast the metric can be computed since for each class the MS-SSIM has to be computed 100*99 times, once for each pair of images. The authors do not discuss how quickly it can actually be done.</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l143">Line 143:</td>
<td colspan="2" class="diff-lineno">Line 143:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second experiment is on the interaction between the Inception accuracy and mean MS-SSIM metric. The author found that they are negatively correlated, and thus that classes that are high quality also tend to be diverse. This is contrary to prevailing wisdom, and since the correlation between them is weak, it appears that it may be only a fluke of the metrics.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second experiment is on the interaction between the Inception accuracy and mean MS-SSIM metric. The author found that they are negatively correlated, and thus that classes that are high quality also tend to be diverse. This is contrary to prevailing wisdom, and since the correlation between them is weak, it appears that it may be only a fluke of the metrics.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The final experiment is on the effect of class splits on image diversity. The authors found that increasing the number of classes handled by each model reduced the diversity of generated images. They make the claim at the beginning of the paper that they show the number of classes is what makes ImageNet synthesis difficult for GANs. This analysis does point in that direction but is not quite conclusive about the issue. Another analysis they could have included towards showing this is how their Inception accuracy metric and the Inception Score are affected by the number of class splits in their model. Perhaps instead of splitting classes among multiple networks, in the future they could augment the classes using more abstract categorical classes as in [[#References|Grinblat et al (2017)]].</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The final experiment is on the effect of class splits on image diversity. The authors found that increasing the number of classes handled by each model reduced the diversity of generated images. They make the claim at the beginning of the paper that they show the number of classes is what makes ImageNet synthesis difficult for GANs. This analysis does point in that direction but is not quite conclusive about the issue. Another analysis they could have included towards showing this is how their Inception accuracy metric and the Inception Score are affected by the number of class splits in their model. Perhaps instead of splitting classes among multiple networks, in the future<ins style="font-weight: bold; text-decoration: none;">, </ins>they could augment the classes using more abstract categorical classes as in [[#References|Grinblat et al (2017)]].</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Conclusion =</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Conclusion =</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l152">Line 152:</td>
<td colspan="2" class="diff-lineno">Line 152:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>In the approach description, the authors use the sum of $L_c + L_s$ as the Loss for $D$; $L_c - L_s$ as the loss for $G$.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>In the approach description, the authors use the sum of $L_c + L_s$ as the Loss for $D$; $L_c - L_s$ as the loss for $G$.</div></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>It would be <del style="font-weight: bold; text-decoration: none;">innteresting </del>if the authors can use a convex combination of them to encourage more real image with the constraint on the class label.</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>It would be <ins style="font-weight: bold; text-decoration: none;">interesting </ins>if the authors can use a convex combination of them to encourage more real image with the constraint on the class label.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= References =</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= References =</div></td></tr>
</table>Aravindbkhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Conditional_Image_Synthesis_with_Auxiliary_Classifier_GANs&diff=31617&oldid=prevAshishgaurav at 17:06, 28 November 20172017-11-28T17:06:09Z<p></p>
<table style="background-color: #fff; color: #202122;" data-mw="interface">
<col class="diff-marker" />
<col class="diff-content" />
<col class="diff-marker" />
<col class="diff-content" />
<tr class="diff-title" lang="us">
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">← Older revision</td>
<td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">Revision as of 13:06, 28 November 2017</td>
</tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l27">Line 27:</td>
<td colspan="2" class="diff-lineno">Line 27:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Formally, let $C\sim p_c$ represent the target class label of each generated observation and $Z$ represent the usual noise vector from the latent space. Then the generator function takes both as arguments to produce image samples: $X_{fake}=G(c,z)$.The discriminator gives a probability distribution over the source $S$ (real or fake) of the image as well as the class label $C$ being generated. $$D(X)={P(S|X),P(C|X)}$$</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Formally, let $C\sim p_c$ represent the target class label of each generated observation and $Z$ represent the usual noise vector from the latent space. Then the generator function takes both as arguments to produce image samples: $X_{fake}=G(c,z)$.The discriminator gives a probability distribution over the source $S$ (real or fake) of the image as well as the class label $C$ being generated. $$D(X)={P(S|X),P(C|X)}$$</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The objective function for the model thus has two parts, one corresponding to the source $L_S$ and the other to the class $L_C$. $D$ is trained to maximize $L_S + L_C$, while $G$ is trained to maximize $L_C-L_S$. Using the notation of Goodfellow et al. (2014), the loss terms are defined as:</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The objective function for the model thus has two parts, one corresponding to the source $L_S$ and the other to the class $L_C$. $D$ is trained to maximize $L_S + L_C$, while $G$ is trained to maximize $L_C-L_S$. Using the notation of <ins style="font-weight: bold; text-decoration: none;">[[#References|</ins>Goodfellow et al. (2014)<ins style="font-weight: bold; text-decoration: none;">]]</ins>, the loss terms are defined as:</div></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>$$L_S=\mathbb{E}_{X\sim p_{data}(x)}[log(P(S=real|X))]+\mathbb{E}_{C,Z\sim p_{C,Z}(c,z)}[log(P(S=fake|G(C,Z)))]$$</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>$$L_S=\mathbb{E}_{X\sim p_{data}(x)}[log(P(S=<ins style="font-weight: bold; text-decoration: none;">\mbox{</ins>real<ins style="font-weight: bold; text-decoration: none;">}</ins>|X))]+\mathbb{E}_{C,Z\sim p_{C,Z}(c,z)}[log(P(S=<ins style="font-weight: bold; text-decoration: none;">\mbox{</ins>fake<ins style="font-weight: bold; text-decoration: none;">}</ins>|G(C,Z)))]$$</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>$$L_C=\mathbb{E}_{X\sim p_{data}(x)}[log(P(C=c|X))]+\mathbb{E}_{C,Z\sim p_{C,Z}(c,z)}[log(P(C=c|G(C,Z)))]$$</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>$$L_C=\mathbb{E}_{X\sim p_{data}(x)}[log(P(C=c|X))]+\mathbb{E}_{C,Z\sim p_{C,Z}(c,z)}[log(P(C=c|G(C,Z)))]$$</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Because G accepts both $C$ and $Z$ as arguments, it is able to learn a mapping $Z\rightarrow X$ that is independent of $C$. The authors argue that all class-specific information should be represented by $C$, allowing $Z$ to represent other factors such as pose, background, etc.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>Because G accepts both $C$ and $Z$ as arguments, it is able to learn a mapping $Z\rightarrow X$ that is independent of $C$. The authors argue that all class-specific information should be represented by $C$, allowing $Z$ to represent other factors such as pose, background, etc.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>Lastly, the authors split the generation process into many class-specific submodels. They point out that the structure of their model permits this split, though it should technically be possible for even the standard GAN framework by dividing the training data into groups according to their known class labels. Other works also employ class splitting with regards to GANs. High level categorical class labels have been shown to improve GAN performance due to the increased abstraction they provide (Grinblat et al. 2017). </div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>Lastly, the authors split the generation process into many class-specific submodels. They point out that the structure of their model permits this split, though it should technically be possible for even the standard GAN framework by dividing the training data into groups according to their known class labels. Other works also employ class splitting with regards to GANs. High level categorical class labels have been shown to improve GAN performance due to the increased abstraction they provide (<ins style="font-weight: bold; text-decoration: none;">[[#References|</ins>Grinblat et al. 2017<ins style="font-weight: bold; text-decoration: none;">]]</ins>). </div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The changes above result in a model capable of generating (some) image samples with both high resolution and global coherence.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The changes above result in a model capable of generating (some) image samples with both high resolution and global coherence.</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l143">Line 143:</td>
<td colspan="2" class="diff-lineno">Line 143:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second experiment is on the interaction between the Inception accuracy and mean MS-SSIM metric. The author found that they are negatively correlated, and thus that classes that are high quality also tend to be diverse. This is contrary to prevailing wisdom, and since the correlation between them is weak, it appears that it may be only a fluke of the metrics.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>The second experiment is on the interaction between the Inception accuracy and mean MS-SSIM metric. The author found that they are negatively correlated, and thus that classes that are high quality also tend to be diverse. This is contrary to prevailing wisdom, and since the correlation between them is weak, it appears that it may be only a fluke of the metrics.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>The final experiment is on the effect of class splits on image diversity. The authors found that increasing the number of classes handled by each model reduced the diversity of generated images. They make the claim at the beginning of the paper that they show the number of classes is what makes ImageNet synthesis difficult for GANs. This analysis does point in that direction but is not quite conclusive about the issue. Another analysis they could have included towards showing this is how their Inception accuracy metric and the Inception Score are affected by the number of class splits in their model. Perhaps instead of splitting classes among multiple networks, in the future they could augment the classes using more abstract categorical classes as in Grinblat et al (2017).</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>The final experiment is on the effect of class splits on image diversity. The authors found that increasing the number of classes handled by each model reduced the diversity of generated images. They make the claim at the beginning of the paper that they show the number of classes is what makes ImageNet synthesis difficult for GANs. This analysis does point in that direction but is not quite conclusive about the issue. Another analysis they could have included towards showing this is how their Inception accuracy metric and the Inception Score are affected by the number of class splits in their model. Perhaps instead of splitting classes among multiple networks, in the future they could augment the classes using more abstract categorical classes as in <ins style="font-weight: bold; text-decoration: none;">[[#References|</ins>Grinblat et al (2017)<ins style="font-weight: bold; text-decoration: none;">]]</ins>.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Conclusion =</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>= Conclusion =</div></td></tr>
<tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l170">Line 170:</td>
<td colspan="2" class="diff-lineno">Line 170:</td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Berg, A. C. (2015). Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 211-252.</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Berg, A. C. (2015). Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 211-252.</div></td></tr>
<tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># G.L. Grinblat, L.C. Uzal, P.M. Granitto. Class-splitting generative adversarial networks. arXiv preprint [https://arxiv.org/abs/1709.07359 : arXiv:1709.07359].</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div># G.L. Grinblat, L.C. Uzal, P.M. Granitto. Class-splitting generative adversarial networks. arXiv preprint [https://arxiv.org/abs/1709.07359 : arXiv:1709.07359].</div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;"></ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">=== Online resources ===</ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">* [https://github.com/buriburisuri/ac-gan github.com/buriburisuri/ac-gan (tensorflow+sugartensor)]</ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">* [https://github.com/kimhc6028/acgan-pytorch github.com/kimhc6028/acgan-pytorch (pytorch)]</ins></div></td></tr>
<tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">* [https://www.youtube.com/watch?v=myP2TN0_MaE Conditional Image Synthesis with Auxiliary Classifier GANs, by Augustus Odena (Video)]</ins></div></td></tr>
</table>Ashishgaurav