Really confused with the leaderboard since as we all can see in the leaderboard are mostly only a few points away from @cdeotte public kernel. I am not saying those who reach this far are all reruns or forks with slight changes and not blaming the kernel since those kernels are pretty good actually.
\n\nSince I am a beginner myself this confuses me since it makes me difficult to benchmark my work (I know I shouldn't put limits to how good my performance is but since I am new I guess I need to know my place). I really would like to know how far a real gan could fare in this competition and the currently available gan models for this competition is very far from the current leaderboard.
\n\nMaybe someone could share a bit their scores with an actual GAN model and possibly hints of which GAN to try.
\n\nTried simple DCGAN implementation so far not so good.
", "messages": [ { "id": "569548", "postDate": "07/06/2019 22:15:00", "content": "Really confused with the leaderboard since as we all can see in the leaderboard are mostly only a few points away from @cdeotte public kernel. I am not saying those who reach this far are all reruns or forks with slight changes and not blaming the kernel since those kernels are pretty good actually.
\n\nSince I am a beginner myself this confuses me since it makes me difficult to benchmark my work (I know I shouldn't put limits to how good my performance is but since I am new I guess I need to know my place). I really would like to know how far a real gan could fare in this competition and the currently available gan models for this competition is very far from the current leaderboard.
\n\nMaybe someone could share a bit their scores with an actual GAN model and possibly hints of which GAN to try.
\n\nTried simple DCGAN implementation so far not so good.
", "rawMarkdown": "Really confused with the leaderboard since as we all can see in the leaderboard are mostly only a few points away from @cdeotte public kernel. I am not saying those who reach this far are all reruns or forks with slight changes and not blaming the kernel since those kernels are pretty good actually. \n\nSince I am a beginner myself this confuses me since it makes me difficult to benchmark my work (I know I shouldn't put limits to how good my performance is but since I am new I guess I need to know my place). I really would like to know how far a real gan could fare in this competition and the currently available gan models for this competition is very far from the current leaderboard.\n\nMaybe someone could share a bit their scores with an actual GAN model and possibly hints of which GAN to try.\n\nTried simple DCGAN implementation so far not so good.", "votes": null }, { "id": "569560", "postDate": "07/06/2019 23:59:37", "content": "My best guess is that
\n\n\n\n\nLB scores 45 - 600 (positions 138 - 314) are conventional GANs\n LB scores 23 - 44 (positions 96 - 137) are blenders from here\n LB scores 15 - 22 (positions 75 - 95) are memorizer GN here\n LB scores 9 - 14 (positions 65 - 74) are blenders from here\n LB scores 5 - 8 (positions 1 - 64) are memorizer GANs here
\n
Blenders have been ruled invalid here. Memorizers GN/GANs are a gray area. In the worse case they output altered training images and are invalid. In the best case, you can turn them into conventional GANs by applying conventional adversarial epochs to an initialized memorizer GAN.
", "rawMarkdown": "My best guess is that \n\n>**LB scores 45 - 600 (positions 138 - 314) are conventional GANs**\nLB scores 23 - 44 (positions 96 - 137) are blenders from [here][1]\nLB scores 15 - 22 (positions 75 - 95) are memorizer GN [here][2]\nLB scores 9 - 14 (positions 65 - 74) are blenders from [here][3]\nLB scores 5 - 8 (positions 1 - 64) are memorizer GANs [here][4]\n\nBlenders have been ruled invalid [here][5]. Memorizers GN/GANs are a gray area. In the worse case they output altered training images and are invalid. In the best case, you can turn them into conventional GANs by applying conventional adversarial epochs to an initialized memorizer GAN.\n\n[1]: https://www.kaggle.com/takumiito/commented-imitation-game\n[2]: https://www.kaggle.com/cdeotte/supervised-generative-dog-net\n[3]: https://www.kaggle.com/jazivxt/imitation-game\n[4]: https://www.kaggle.com/cdeotte/dog-memorizer-gan\n[5]: https://www.kaggle.com/c/generative-dog-images/discussion/98183", "votes": null }, { "id": "571107", "postDate": "07/09/2019 07:03:46", "content": "Don't worry, in this competition everyone is confused with the leaderboard :)
", "rawMarkdown": "Don't worry, in this competition everyone is confused with the leaderboard :)", "votes": null } ], "comments": [ { "id": 569560, "author_name": "cdeotte", "author_url": "", "post_date": "07/06/2019 23:59:37", "content": "My best guess is that
\n\n\n\n\nLB scores 45 - 600 (positions 138 - 314) are conventional GANs\n LB scores 23 - 44 (positions 96 - 137) are blenders from here\n LB scores 15 - 22 (positions 75 - 95) are memorizer GN here\n LB scores 9 - 14 (positions 65 - 74) are blenders from here\n LB scores 5 - 8 (positions 1 - 64) are memorizer GANs here
\n
Blenders have been ruled invalid here. Memorizers GN/GANs are a gray area. In the worse case they output altered training images and are invalid. In the best case, you can turn them into conventional GANs by applying conventional adversarial epochs to an initialized memorizer GAN.
", "votes": null, "replies": [] }, { "id": 571107, "author_name": "demonplus", "author_url": "", "post_date": "07/09/2019 07:03:46", "content": "Don't worry, in this competition everyone is confused with the leaderboard :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "569548": "Really confused with the leaderboard since as we all can see in the leaderboard are mostly only a few points away from @cdeotte public kernel. I am not saying those who reach this far are all reruns or forks with slight changes and not blaming the kernel since those kernels are pretty good actually. \n\nSince I am a beginner myself this confuses me since it makes me difficult to benchmark my work (I know I shouldn't put limits to how good my performance is but since I am new I guess I need to know my place). I really would like to know how far a real gan could fare in this competition and the currently available gan models for this competition is very far from the current leaderboard.\n\nMaybe someone could share a bit their scores with an actual GAN model and possibly hints of which GAN to try.\n\nTried simple DCGAN implementation so far not so good.", "569560": "My best guess is that \n\n>**LB scores 45 - 600 (positions 138 - 314) are conventional GANs**\nLB scores 23 - 44 (positions 96 - 137) are blenders from [here][1]\nLB scores 15 - 22 (positions 75 - 95) are memorizer GN [here][2]\nLB scores 9 - 14 (positions 65 - 74) are blenders from [here][3]\nLB scores 5 - 8 (positions 1 - 64) are memorizer GANs [here][4]\n\nBlenders have been ruled invalid [here][5]. Memorizers GN/GANs are a gray area. In the worse case they output altered training images and are invalid. In the best case, you can turn them into conventional GANs by applying conventional adversarial epochs to an initialized memorizer GAN.\n\n[1]: https://www.kaggle.com/takumiito/commented-imitation-game\n[2]: https://www.kaggle.com/cdeotte/supervised-generative-dog-net\n[3]: https://www.kaggle.com/jazivxt/imitation-game\n[4]: https://www.kaggle.com/cdeotte/dog-memorizer-gan\n[5]: https://www.kaggle.com/c/generative-dog-images/discussion/98183", "571107": "Don't worry, in this competition everyone is confused with the leaderboard :)" }, "source": "meta" }