![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2779868%2F351e45b4fa40540839a5899dd9c02184%2FScreenshot%20from%202019-08-14%2023-52-23.png?generation=1565819632119582&alt=media\")
1st of all... good luck to all of you :)
\n\nBigGan\n- Dmitry Vorobiev[14.95516]\n- Big GAN 256 - [LB 15] by @cdeotte \n- @bestfitting also used this technique.
\n\nDCGAN\n- Johannl[38.26191]\n- Happy GAN[38.70295]\n- hirune924[38.92350]
\n\nSNGAN + projection\n- My approach (gold?)
\n\nPlease, click and read the paper if you want to learn more about GANs
\n\n> MNIST. We validate our claim (from Section 3) that a GAN can perfectlyreproduce a simple dataset on MNIST. A four-layer convnet classifier trained onreal MNIST data achieves 99.3% accuracy on the test set. In contrast, imagesgenerated with SNGAN achieve a GAN-train accuracy of 99.0% and GAN-testaccuracy of 99.2%, highlighting their high image quality as well as diversity.
\n\n> CIFAR10. Table 1 shows a comparison of state-of-the-art GAN models onCIFAR10. We observe that the relative ranking of models is consistent acrossdifferent metrics: FID, GAN-train and GAN-test accuracies. Both GAN-trainand GAN-test are quite high for SNGAN and WGAN-GP (10M). This impliesthat both the image quality and the diversity are good, but are still lower thanthat of real images (92.8 in the first row). Note that PixelCNN++ has lowdiversity because GAN-test is much higher than GAN-train in this case. This isin line with its relatively poor Inception score and FID (as shown in [32] FID isquite sensitive to mode dropping).
\n\n
> ImageNet. On this dataset, which is one of the more challenging ones for imagesynthesis [36], we analyzed the performance of the two best GAN models basedon our CIFAR experiments, i.e., SNGAN and WGAN-GP. As shown in Table 3,SNGAN achieves a reasonable GAN-train accuracy and a relatively high GAN-test accuracy at 128×128 resolution. This suggests that SNGAN generatedimages have good quality, but their diversity is much lower than the original data.
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2779868%2F824cd56791601348967aa84e91807c5d%2FScreenshot%20from%202019-08-14%2023-52-41.png?generation=1565819636563070&alt=media\")
> Reading this paper allowed me to know more about SNGAN and how it outperforms. I guess more peoples realized about this and that's why they didn't shared it (kernel or discussions). Also WGAN, but in this case it worked really bad. And the results from these models are really important in order to measure my performance in this comp because the best GANs on ImageNet score around 30 FID :)
\n\nAs you can read here: If we have to give away prize based on public LB submissions right now the best results are also around 20-30 ;)
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2779868%2F8a2e628fb18527f08477c8d013b5fa2f%2FScreenshot%20from%202019-08-14%2023-53-06.png?generation=1565819634909044&alt=media\")
1st of all... good luck to all of you :)
\n\nBigGan\n- Dmitry Vorobiev[14.95516]\n- Big GAN 256 - [LB 15] by @cdeotte \n- @bestfitting also used this technique.
\n\nDCGAN\n- Johannl[38.26191]\n- Happy GAN[38.70295]\n- hirune924[38.92350]
\n\nSNGAN + projection\n- My approach (gold?)
\n\nPlease, click and read the paper if you want to learn more about GANs
\n\n> MNIST. We validate our claim (from Section 3) that a GAN can perfectlyreproduce a simple dataset on MNIST. A four-layer convnet classifier trained onreal MNIST data achieves 99.3% accuracy on the test set. In contrast, imagesgenerated with SNGAN achieve a GAN-train accuracy of 99.0% and GAN-testaccuracy of 99.2%, highlighting their high image quality as well as diversity.
\n\n> CIFAR10. Table 1 shows a comparison of state-of-the-art GAN models onCIFAR10. We observe that the relative ranking of models is consistent acrossdifferent metrics: FID, GAN-train and GAN-test accuracies. Both GAN-trainand GAN-test are quite high for SNGAN and WGAN-GP (10M). This impliesthat both the image quality and the diversity are good, but are still lower thanthat of real images (92.8 in the first row). Note that PixelCNN++ has lowdiversity because GAN-test is much higher than GAN-train in this case. This isin line with its relatively poor Inception score and FID (as shown in [32] FID isquite sensitive to mode dropping).
\n\n
> ImageNet. On this dataset, which is one of the more challenging ones for imagesynthesis [36], we analyzed the performance of the two best GAN models basedon our CIFAR experiments, i.e., SNGAN and WGAN-GP. As shown in Table 3,SNGAN achieves a reasonable GAN-train accuracy and a relatively high GAN-test accuracy at 128×128 resolution. This suggests that SNGAN generatedimages have good quality, but their diversity is much lower than the original data.
\n\n
> Reading this paper allowed me to know more about SNGAN and how it outperforms. I guess more peoples realized about this and that's why they didn't shared it (kernel or discussions). Also WGAN, but in this case it worked really bad. And the results from these models are really important in order to measure my performance in this comp because the best GANs on ImageNet score around 30 FID :)
\n\nAs you can read here: If we have to give away prize based on public LB submissions right now the best results are also around 20-30 ;)
\n\n
Thank you for sharing Nanashi..!!!\nBig help ...!!
", "rawMarkdown": "Thank you for sharing Nanashi..!!!\nBig help ...!!", "votes": null } ], "comments": [ { "id": 599672, "author_name": "damnboy", "author_url": "", "post_date": "08/15/2019 08:36:37", "content": "Thank you for sharing Nanashi..!!!\nBig help ...!!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "599374": "**1st of all... good luck to all of you :)**\n\n----\n\n### Best Solutions\n\n**BigGan**\n- [Dmitry Vorobiev[14.95516]](https://www.kaggle.com/dvorobiev/doggies-biggan-sub-data-aug-3)\n- [Big GAN 256 - [LB 15]](https://www.kaggle.com/cdeotte/big-gan-256-lb-15) by @cdeotte \n- @bestfitting also used this technique.\n\n**DCGAN**\n- [Johannl[38.26191]](https://www.kaggle.com/johannl/sn-dcgan-style?scriptVersionId=18780768)\n- [Happy GAN[38.70295]](https://www.kaggle.com/leonshangguan/dcgan-data-cleaning-sub-v1?scriptVersionId=18470145)\n- [hirune924[38.92350]](https://www.kaggle.com/hirune924/public-lb-38-92350-solution)\n\n**SNGAN + projection**\n- [My approach (gold?)](https://www.kaggle.com/c/generative-dog-images/discussion/104170#latest-599602)\n\n----\n\n### [How good is my GAN?](https://arxiv.org/abs/1807.09499)\n Please, click and read the paper if you want to learn more about GANs\n\n> **MNIST.** We validate our claim (from Section 3) that a GAN can perfectlyreproduce a simple dataset on MNIST. A four-layer convnet classifier trained onreal MNIST data achieves 99.3% accuracy on the test set. In contrast, imagesgenerated with SNGAN achieve a GAN-train accuracy of 99.0% and GAN-testaccuracy of 99.2%, highlighting their high image quality as well as diversity.\n\n---\n\n> **CIFAR10.** Table 1 shows a comparison of state-of-the-art GAN models onCIFAR10. We observe that the relative ranking of models is consistent acrossdifferent metrics: FID, GAN-train and GAN-test accuracies. Both GAN-trainand GAN-test are quite high for SNGAN and WGAN-GP (10M). This impliesthat both the image quality and the diversity are good, but are still lower thanthat of real images (92.8 in the first row). Note that PixelCNN++ has lowdiversity because GAN-test is much higher than GAN-train in this case. This isin line with its relatively poor Inception score and FID (as shown in [32] FID isquite sensitive to mode dropping).\n\n\n\n---\n\n> **ImageNet.** On this dataset, which is one of the more challenging ones for imagesynthesis [36], we analyzed the performance of the two best GAN models basedon our CIFAR experiments, i.e., SNGAN and WGAN-GP. As shown in Table 3,SNGAN achieves a reasonable GAN-train accuracy and a relatively high GAN-test accuracy at 128×128 resolution. This suggests that SNGAN generatedimages have good quality, but their diversity is much lower than the original data.\n\n\n\n\n> Reading this paper allowed me to know more about **SNGAN** and how it outperforms. I guess more peoples realized about this and that's why they didn't shared it (kernel or discussions). Also **WGAN**, but in this case it worked really bad. And the results from these models are really important in order to measure my performance in this comp because the best GANs on **ImageNet** score around 30 FID :)\n\nAs you can read here: [If we have to give away prize based on public LB submissions right now](https://www.kaggle.com/c/generative-dog-images/discussion/104028#latest-599258) the best results are also around 20-30 ;)\n\n", "599672": "Thank you for sharing Nanashi..!!!\nBig help ...!!" }, "source": "meta" }