I'm a noob here, I've only participated in about 18 competitions, but this seems very strange to me.\n- I don't know what's allowed and what's not.\n- I don't know if other people are using legal methods or not.\n- LB is broken. All this scores are from \"ilegal\" kernels:\n23.93306, 24.39227, 34.13305, 34.13304 etc\n- I'm not even absolutely sure what I'm doing is right.\n- I'm not 100% sure that @cdeotte methods are allowed. Dog Autoencoder is not allowed, and I don't know if this wonderful kernel Supervised Generative Dog Net is allowed either.
> ... your trained generator model should be able to stand alone and produce images without the input of training images. ...
\n\nWhat do you think? Maybe I'm a whiner or a paranoid, I don't understand what's going on, I really like this topic, but I wouldn't want to lose a month here.
\n\n
", "messages": [ { "id": 567721, "postDate": "2019-07-03T22:48:36.643Z", "content": "I'm a noob here, I've only participated in about 18 competitions, but this seems very strange to me.\n- I don't know what's allowed and what's not.\n- I don't know if other people are using legal methods or not.\n- LB is broken. All this scores are from \"ilegal\" kernels:\n23.93306, 24.39227, 34.13305, 34.13304 etc\n- I'm not even absolutely sure what I'm doing is right.\n- I'm not 100% sure that @cdeotte methods are allowed. Dog Autoencoder is not allowed, and I don't know if this wonderful kernel Supervised Generative Dog Net is allowed either.
> ... your trained generator model should be able to stand alone and produce images without the input of training images. ...
\n\nWhat do you think? Maybe I'm a whiner or a paranoid, I don't understand what's going on, I really like this topic, but I wouldn't want to lose a month here.
\n\n
", "rawMarkdown": "I'm a **noob** here, I've only participated in about 18 competitions, but this seems very strange to me.\n- I don't know what's allowed and what's not.\n- I don't know if other people are using legal methods or not.\n- LB is broken. All this scores are from *\"ilegal\"* kernels:\n```23.93306, 24.39227, 34.13305, 34.13304 etc```\n- I'm not even absolutely sure what I'm doing is right.\n- I'm not 100% sure that @cdeotte methods are allowed. [Dog Autoencoder](https://www.kaggle.com/cdeotte/dog-autoencoder) is not allowed, and I don't know if this wonderful kernel [Supervised Generative Dog Net](https://www.kaggle.com/cdeotte/supervised-generative-dog-net) is allowed either.\n\n> ... your trained generator model should be able to stand alone and produce images without the input of training images. ...\n\n- FID is easily fooled, pretty sure that kaggle won't check kernels one by one, even then ... what is a GAN? there are tons of GANs... how do you prove your method is a GAN or VAE or something allowed? the **grey area** is massive... 😄 😄\n\nWhat do you think? Maybe I'm a whiner or a paranoid, I don't understand what's going on, I really like this topic, but I wouldn't want to lose a month here.\n\n", "votes": 27 }, { "id": 567837, "postDate": "2019-07-04T04:24:54.680Z", "content": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯
", "rawMarkdown": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯", "votes": 12 }, { "id": 568355, "postDate": "2019-07-04T19:07:43.590Z", "content": "There is one person who could be a benchmark for you - it is yourself\nYou try to beat yourself, become better and hope it will end somewhere high in the private
", "rawMarkdown": "There is one person who could be a benchmark for you - it is yourself\nYou try to beat yourself, become better and hope it will end somewhere high in the private", "votes": 5, "replies": [ { "id": 580759, "postDate": "2019-07-20T17:35:34.910Z", "content": "That is a really nice way to put it. :) The best indicator is your personal performance. The \"you\" of today vs yesterday's you.
", "rawMarkdown": "That is a really nice way to put it. :) The best indicator is your personal performance. The \"you\" of today vs yesterday's you. ", "votes": 1 } ] }, { "id": 571606, "postDate": "2019-07-09T20:06:31.520Z", "content": "More and more I become convinced the best Kaggle strategy is to join competitions at the last 2 weeks.
", "rawMarkdown": "More and more I become convinced the best Kaggle strategy is to join competitions at the last 2 weeks.", "votes": 3 }, { "id": 578501, "postDate": "2019-07-17T19:34:48.063Z", "content": "\n\n\nI'm a noob here, I've only participated in about 18 competitions
\n
Did I miss something? That seems like a pro to me.
", "rawMarkdown": "> I'm a noob here, I've only participated in about 18 competitions\n\nDid I miss something? That seems like a pro to me. ", "votes": 1, "replies": [ { "id": 580758, "postDate": "2019-07-20T17:33:05.680Z", "content": "I guess that was just to increase the level of drama. This is barely my first competition and I can't even imagine participating in 18. Nice to meet everyone!
", "rawMarkdown": "I guess that was just to increase the level of drama. This is barely my first competition and I can't even imagine participating in 18. Nice to meet everyone!", "votes": 3 } ] }, { "id": 571129, "postDate": "2019-07-09T07:44:04.237Z", "content": "Has anyone reached a single digit MFid with pure GANs and methods that are allowed?\nIt's really hard to comprehend the lb\nEDIT: I am also confused over what even is a \"pure GAN\" chris' kernels, while being really informative, confused the hell out of me regarding the rules of this competition.
", "rawMarkdown": "Has anyone reached a single digit MFid with pure GANs and methods that are allowed?\nIt's really hard to comprehend the lb\nEDIT: I am also confused over what even is a \"pure GAN\" chris' kernels, while being really informative, confused the hell out of me regarding the rules of this competition. ", "votes": 1, "replies": [ { "id": 571139, "postDate": "2019-07-09T08:05:26.737Z", "content": "I’ve fed genuine good quality non-dataset dog images into the metric and it scored around 20. It’s hard to imagine getting into single digits.
\n\nThe rules are pretty simple. You can’t take your homework into the exam. You have to generate the images from scratch.
\n\nSo you can’t directly use training images to create new images, eg style transfer. You can’t train a gan as a memoriser, because that’s just using the training images. However, Chris brilliantly shows that all GANs suffer from some memorisation, by accident, if not design. Authors and evaluators may be oblivious to the degree of memorisation, and it is difficult to ‘detect’.
\n\nYou are also not allowed to use external data. Clearly it means datasets and trained models. And obviously you shouldn’t encode output pngs in a notebook. But there are gray areas here too that the forum hasn’t explored much.
", "rawMarkdown": "I’ve fed genuine good quality non-dataset dog images into the metric and it scored around 20. It’s hard to imagine getting into single digits. \n\nThe rules are pretty simple. You can’t take your homework into the exam. You have to generate the images from scratch. \n\nSo you can’t directly use training images to create new images, eg style transfer. You can’t train a gan as a memoriser, because that’s just using the training images. However, Chris brilliantly shows that all GANs suffer from some memorisation, by accident, if not design. Authors and evaluators may be oblivious to the degree of memorisation, and it is difficult to ‘detect’. \n\nYou are also not allowed to use external data. Clearly it means datasets and trained models. And obviously you shouldn’t encode output pngs in a notebook. But there are gray areas here too that the forum hasn’t explored much.", "votes": 4 }, { "id": 571156, "postDate": "2019-07-09T08:46:48.620Z", "content": "Thank you for the clarification
", "rawMarkdown": "Thank you for the clarification" } ] }, { "id": 571112, "postDate": "2019-07-09T07:14:10.077Z", "content": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯
", "rawMarkdown": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯", "votes": 1, "replies": [ { "id": 571465, "postDate": "2019-07-09T16:02:20.170Z", "content": "Never ;) still working hard\nOh, and know I'm an alchemist 😄
", "rawMarkdown": "Never ;) still working hard\nOh, and know I'm an alchemist 😄 ", "votes": 2 }, { "id": 571866, "postDate": "2019-07-10T07:00:51.757Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 568275, "postDate": "2019-07-04T16:20:14.717Z", "content": "This is my second competetion . I don't know about leaderboards as my skill is not at that level to expect anything \nout of this competetion. But your kernel was the first thing I read to understand what is going on and how to approach the problem. This won't probably give you a material benefit . But just so you know , there is a larger value that you give back and how much people like me appreciates it . \nWould love if there is a reconciliation done mid competetion and scoreboard resets . Would be happy for you to jump up in LB for your amazing work .
", "rawMarkdown": "This is my second competetion . I don't know about leaderboards as my skill is not at that level to expect anything \nout of this competetion. But your kernel was the first thing I read to understand what is going on and how to approach the problem. This won't probably give you a material benefit . But just so you know , there is a larger value that you give back and how much people like me appreciates it . \nWould love if there is a reconciliation done mid competetion and scoreboard resets . Would be happy for you to jump up in LB for your amazing work .", "votes": 1 }, { "id": 568242, "postDate": "2019-07-04T15:29:45.957Z", "content": "I just have the opportunity to fulfill my long lasting wish of generating good doggos, which is dope, and I wouldn't have it any other way. <3
\n\nOther than that, I agree with you that in a gray zone which is uncomfortable for a \"competition\".
\n\nDon't you think https://www.kaggle.com/c/generative-dog-images/discussion/98229 is a solution ? Can't kaggle check the top 10 private kernels ? I think we can reasonably assume that this competition will not crumble just because of the incapacity of MIFID to detect non-GAN solutions.
\n\n> what is a GAN? there are tons of GANs… how do you prove your method is a GAN or VAE or something allowed? the grey area is massive… 😄 😄
\n\nToo philosophical lel... however we could try to set a definition for this competition.
\n\nFrom my understanding the \"adversarial\" usually refers to the fact that GAN are two agents competing, one for generation and the other for discrimination. The \"net\" means that you have to use an ANN.
\n\nThis means that https://www.kaggle.com/cdeotte/supervised-generative-dog-net is not eligible because it does learn a latent space but it does no adversarial rounds in a \"game theory\" manner.
\n\nThat's both just my noob opinion here and my personal attempt to find a way out of this mess 😄
", "rawMarkdown": "I just have the opportunity to fulfill my long lasting wish of generating **good doggos**, which is dope, and I wouldn't have it any other way. <3\n\nOther than that, I agree with you that in a gray zone which is uncomfortable for a \"competition\".\n\nDon't you think https://www.kaggle.com/c/generative-dog-images/discussion/98229 is a solution ? Can't kaggle check the top 10 private kernels ? I think we can reasonably assume that this competition will not crumble just because of the incapacity of MIFID to detect non-GAN solutions.\n\n> what is a GAN? there are tons of GANs… how do you prove your method is a GAN or VAE or something allowed? the grey area is massive… 😄 😄\n\nToo philosophical lel... however we could try to set a definition *for this competition*.\n\nFrom my understanding the \"adversarial\" usually refers to the fact that GAN are two agents competing, one for generation and the other for discrimination. The \"net\" means that you have to use an ANN.\n\nThis means that https://www.kaggle.com/cdeotte/supervised-generative-dog-net is not eligible because it does learn a latent space but it does no adversarial rounds in a \"game theory\" manner.\n\nThat's both just my noob opinion here and my personal attempt to find a way out of this mess 😄 ", "votes": 2, "replies": [ { "id": 568253, "postDate": "2019-07-04T15:46:42.410Z", "content": "What I think that makes Nanashi, and almost everyone else, uneasy is the idea that if someone is not in the top-10 she/he will not be checked. Basically it is the same as saying you can cheat as much as you want in this competition unless you risk ending up in the top-10...a lot of Kagglers are telling me they intend to stay away from this competition for this reason
", "rawMarkdown": "What I think that makes Nanashi, and almost everyone else, uneasy is the idea that if someone is not in the top-10 she/he will not be checked. Basically it is the same as saying you can cheat as much as you want in this competition unless you risk ending up in the top-10...a lot of Kagglers are telling me they intend to stay away from this competition for this reason", "votes": 5 }, { "id": 568305, "postDate": "2019-07-04T17:18:31.053Z", "content": "Makes sense...
\n\nHope the Kaggle team will find a way to fix this issue. 😢
", "rawMarkdown": "Makes sense...\n\nHope the Kaggle team will find a way to fix this issue. 😢 ", "votes": 1 } ] }, { "id": 568577, "postDate": "2019-07-05T06:26:31.330Z", "content": "If upvote could cheer you up, I would upvote you 3000 times. Your kernels are great, your one month are surely contributing to many people's knowledge.
", "rawMarkdown": "If upvote could cheer you up, I would upvote you 3000 times. Your kernels are great, your one month are surely contributing to many people's knowledge. " }, { "id": 568464, "postDate": "2019-07-05T01:02:27.050Z", "content": "Agree with generative models should be models that could work in inference time with no target information. The LB is ruined, so sorry for that. BTW, how to finetuning the parameters for gan..... I have played with gan before, but only want to see them work and never try to improve the performance to be best.
", "rawMarkdown": "Agree with generative models should be models that could work in inference time with no target information. The LB is ruined, so sorry for that. BTW, how to finetuning the parameters for gan..... I have played with gan before, but only want to see them work and never try to improve the performance to be best." } ], "comments": [ { "id": 567837, "author_name": "jazivxt", "author_url": "", "post_date": "2019-07-04T04:24:54.680000", "content": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯
", "votes": 12, "replies": [] }, { "id": 568355, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2019-07-04T19:07:43.590000", "content": "There is one person who could be a benchmark for you - it is yourself\nYou try to beat yourself, become better and hope it will end somewhere high in the private
", "votes": 5, "replies": [ { "id": 580759, "author_name": "Geosphere", "author_url": "", "post_date": "2019-07-20T17:35:34.910000", "content": "That is a really nice way to put it. :) The best indicator is your personal performance. The \"you\" of today vs yesterday's you.
", "votes": 1, "replies": [] } ] }, { "id": 571606, "author_name": "Mohammed Tayor", "author_url": "", "post_date": "2019-07-09T20:06:31.520000", "content": "More and more I become convinced the best Kaggle strategy is to join competitions at the last 2 weeks.
", "votes": 3, "replies": [] }, { "id": 578501, "author_name": "ryches", "author_url": "", "post_date": "2019-07-17T19:34:48.063000", "content": "\n\n\nI'm a noob here, I've only participated in about 18 competitions
\n
Did I miss something? That seems like a pro to me.
", "votes": 1, "replies": [ { "id": 580758, "author_name": "Geosphere", "author_url": "", "post_date": "2019-07-20T17:33:05.680000", "content": "I guess that was just to increase the level of drama. This is barely my first competition and I can't even imagine participating in 18. Nice to meet everyone!
", "votes": 3, "replies": [] } ] }, { "id": 571129, "author_name": "timetraveller", "author_url": "", "post_date": "2019-07-09T07:44:04.237000", "content": "Has anyone reached a single digit MFid with pure GANs and methods that are allowed?\nIt's really hard to comprehend the lb\nEDIT: I am also confused over what even is a \"pure GAN\" chris' kernels, while being really informative, confused the hell out of me regarding the rules of this competition.
", "votes": 1, "replies": [ { "id": 571139, "author_name": "robga", "author_url": "", "post_date": "2019-07-09T08:05:26.737000", "content": "I’ve fed genuine good quality non-dataset dog images into the metric and it scored around 20. It’s hard to imagine getting into single digits.
\n\nThe rules are pretty simple. You can’t take your homework into the exam. You have to generate the images from scratch.
\n\nSo you can’t directly use training images to create new images, eg style transfer. You can’t train a gan as a memoriser, because that’s just using the training images. However, Chris brilliantly shows that all GANs suffer from some memorisation, by accident, if not design. Authors and evaluators may be oblivious to the degree of memorisation, and it is difficult to ‘detect’.
\n\nYou are also not allowed to use external data. Clearly it means datasets and trained models. And obviously you shouldn’t encode output pngs in a notebook. But there are gray areas here too that the forum hasn’t explored much.
", "votes": 4, "replies": [] }, { "id": 571156, "author_name": "timetraveller", "author_url": "", "post_date": "2019-07-09T08:46:48.620000", "content": "Thank you for the clarification
", "votes": 0, "replies": [] } ] }, { "id": 571112, "author_name": "jazivxt", "author_url": "", "post_date": "2019-07-09T07:14:10.077000", "content": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯
", "votes": 1, "replies": [ { "id": 571465, "author_name": "Nanashi", "author_url": "", "post_date": "2019-07-09T16:02:20.170000", "content": "Never ;) still working hard\nOh, and know I'm an alchemist 😄
", "votes": 2, "replies": [] }, { "id": 571866, "author_name": "", "author_url": "", "post_date": "2019-07-10T07:00:51.757000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 568275, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-04T16:20:14.717000", "content": "This is my second competetion . I don't know about leaderboards as my skill is not at that level to expect anything \nout of this competetion. But your kernel was the first thing I read to understand what is going on and how to approach the problem. This won't probably give you a material benefit . But just so you know , there is a larger value that you give back and how much people like me appreciates it . \nWould love if there is a reconciliation done mid competetion and scoreboard resets . Would be happy for you to jump up in LB for your amazing work .
", "votes": 1, "replies": [] }, { "id": 568242, "author_name": "Guillaume Desforges", "author_url": "", "post_date": "2019-07-04T15:29:45.957000", "content": "I just have the opportunity to fulfill my long lasting wish of generating good doggos, which is dope, and I wouldn't have it any other way. <3
\n\nOther than that, I agree with you that in a gray zone which is uncomfortable for a \"competition\".
\n\nDon't you think https://www.kaggle.com/c/generative-dog-images/discussion/98229 is a solution ? Can't kaggle check the top 10 private kernels ? I think we can reasonably assume that this competition will not crumble just because of the incapacity of MIFID to detect non-GAN solutions.
\n\n> what is a GAN? there are tons of GANs… how do you prove your method is a GAN or VAE or something allowed? the grey area is massive… 😄 😄
\n\nToo philosophical lel... however we could try to set a definition for this competition.
\n\nFrom my understanding the \"adversarial\" usually refers to the fact that GAN are two agents competing, one for generation and the other for discrimination. The \"net\" means that you have to use an ANN.
\n\nThis means that https://www.kaggle.com/cdeotte/supervised-generative-dog-net is not eligible because it does learn a latent space but it does no adversarial rounds in a \"game theory\" manner.
\n\nThat's both just my noob opinion here and my personal attempt to find a way out of this mess 😄
", "votes": 2, "replies": [ { "id": 568253, "author_name": "Luca Massaron", "author_url": "", "post_date": "2019-07-04T15:46:42.410000", "content": "What I think that makes Nanashi, and almost everyone else, uneasy is the idea that if someone is not in the top-10 she/he will not be checked. Basically it is the same as saying you can cheat as much as you want in this competition unless you risk ending up in the top-10...a lot of Kagglers are telling me they intend to stay away from this competition for this reason
", "votes": 5, "replies": [] }, { "id": 568305, "author_name": "Guillaume Desforges", "author_url": "", "post_date": "2019-07-04T17:18:31.053000", "content": "Makes sense...
\n\nHope the Kaggle team will find a way to fix this issue. 😢
", "votes": 1, "replies": [] } ] }, { "id": 568577, "author_name": "yukiya", "author_url": "", "post_date": "2019-07-05T06:26:31.330000", "content": "If upvote could cheer you up, I would upvote you 3000 times. Your kernels are great, your one month are surely contributing to many people's knowledge.
", "votes": 0, "replies": [] }, { "id": 568464, "author_name": "jionie", "author_url": "", "post_date": "2019-07-05T01:02:27.050000", "content": "Agree with generative models should be models that could work in inference time with no target information. The LB is ruined, so sorry for that. BTW, how to finetuning the parameters for gan..... I have played with gan before, but only want to see them work and never try to improve the performance to be best.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "567721": "I'm a **noob** here, I've only participated in about 18 competitions, but this seems very strange to me.\n- I don't know what's allowed and what's not.\n- I don't know if other people are using legal methods or not.\n- LB is broken. All this scores are from *\"ilegal\"* kernels:\n```23.93306, 24.39227, 34.13305, 34.13304 etc```\n- I'm not even absolutely sure what I'm doing is right.\n- I'm not 100% sure that @cdeotte methods are allowed. [Dog Autoencoder](https://www.kaggle.com/cdeotte/dog-autoencoder) is not allowed, and I don't know if this wonderful kernel [Supervised Generative Dog Net](https://www.kaggle.com/cdeotte/supervised-generative-dog-net) is allowed either.\n\n> ... your trained generator model should be able to stand alone and produce images without the input of training images. ...\n\n- FID is easily fooled, pretty sure that kaggle won't check kernels one by one, even then ... what is a GAN? there are tons of GANs... how do you prove your method is a GAN or VAE or something allowed? the **grey area** is massive... 😄 😄\n\nWhat do you think? Maybe I'm a whiner or a paranoid, I don't understand what's going on, I really like this topic, but I wouldn't want to lose a month here.\n\n", "567837": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯", "568355": "There is one person who could be a benchmark for you - it is yourself\nYou try to beat yourself, become better and hope it will end somewhere high in the private", "571606": "More and more I become convinced the best Kaggle strategy is to join competitions at the last 2 weeks.", "578501": "> I'm a noob here, I've only participated in about 18 competitions\n\nDid I miss something? That seems like a pro to me. ", "571129": "Has anyone reached a single digit MFid with pure GANs and methods that are allowed?\nIt's really hard to comprehend the lb\nEDIT: I am also confused over what even is a \"pure GAN\" chris' kernels, while being really informative, confused the hell out of me regarding the rules of this competition. ", "571112": "Ｈ𝐀𝑷𝑷𝓎 🇰𝗮𝘨𝘨🇱𝖎Ｎɢ 💯", "568275": "This is my second competetion . I don't know about leaderboards as my skill is not at that level to expect anything \nout of this competetion. But your kernel was the first thing I read to understand what is going on and how to approach the problem. This won't probably give you a material benefit . But just so you know , there is a larger value that you give back and how much people like me appreciates it . \nWould love if there is a reconciliation done mid competetion and scoreboard resets . Would be happy for you to jump up in LB for your amazing work .", "568242": "I just have the opportunity to fulfill my long lasting wish of generating **good doggos**, which is dope, and I wouldn't have it any other way. <3\n\nOther than that, I agree with you that in a gray zone which is uncomfortable for a \"competition\".\n\nDon't you think https://www.kaggle.com/c/generative-dog-images/discussion/98229 is a solution ? Can't kaggle check the top 10 private kernels ? I think we can reasonably assume that this competition will not crumble just because of the incapacity of MIFID to detect non-GAN solutions.\n\n> what is a GAN? there are tons of GANs… how do you prove your method is a GAN or VAE or something allowed? the grey area is massive… 😄 😄\n\nToo philosophical lel... however we could try to set a definition *for this competition*.\n\nFrom my understanding the \"adversarial\" usually refers to the fact that GAN are two agents competing, one for generation and the other for discrimination. The \"net\" means that you have to use an ANN.\n\nThis means that https://www.kaggle.com/cdeotte/supervised-generative-dog-net is not eligible because it does learn a latent space but it does no adversarial rounds in a \"game theory\" manner.\n\nThat's both just my noob opinion here and my personal attempt to find a way out of this mess 😄 ", "568577": "If upvote could cheer you up, I would upvote you 3000 times. Your kernels are great, your one month are surely contributing to many people's knowledge. ", "568464": "Agree with generative models should be models that could work in inference time with no target information. The LB is ruined, so sorry for that. BTW, how to finetuning the parameters for gan..... I have played with gan before, but only want to see them work and never try to improve the performance to be best." } }![https://media1.tenor.com/images/357a6e777b182ba17f679483d5d4a676/tenor.gif?itemid=11760248](\"https://media1.tenor.com/images/357a6e777b182ba17f679483d5d4a676/tenor.gif?itemid=11760248\"){kind=link}