I want to start working on this competition but before i invest my time in this one, I would really like to know what is Kaggle planning to do in the end.
\n\nI really don’t want to end up wasting my time because of a bunch of people who didn’t follow the rules.
", "messages": [ { "id": "579825", "postDate": "07/19/2019 09:11:27", "content": "I want to start working on this competition but before i invest my time in this one, I would really like to know what is Kaggle planning to do in the end.
\n\nI really don’t want to end up wasting my time because of a bunch of people who didn’t follow the rules.
", "rawMarkdown": "I want to start working on this competition but before i invest my time in this one, I would really like to know what is Kaggle planning to do in the end.\n\n1. Will all the submissions not using a GAN approach be removed from LB?\n2. Will only the winner’s solution be verified?\n3. If the winners choose not to disclose their solution for forfeit prize money, will they be removed from LB?\n4. How many ranks will be checked for GAN based approach?\n\nI really don’t want to end up wasting my time because of a bunch of people who didn’t follow the rules.", "votes": null }, { "id": "579911", "postDate": "07/19/2019 11:45:50", "content": "Great question. There's also the fear of being incorrectly disqualified. If you invent a new kind of GAN that beats all published research paper GANs, it may get disqualified because the images look so good that Kaggle thinks you are memorizing training images (even if you are actually generalizing).
", "rawMarkdown": "Great question. There's also the fear of being incorrectly disqualified. If you invent a new kind of GAN that beats all published research paper GANs, it may get disqualified because the images look so good that Kaggle thinks you are memorizing training images (even if you are actually generalizing).", "votes": null }, { "id": "579915", "postDate": "07/19/2019 11:53:43", "content": "One can easily prove that and come back on the LB.
", "rawMarkdown": "One can easily prove that and come back on the LB.", "votes": null }, { "id": "579926", "postDate": "07/19/2019 12:03:43", "content": "Unfortunately there is no successful metric that distinguishes memorization from generalization which makes proving anything difficult.
\n\nWhat if a GAN memorizes all the images first. And then it's put through a process that converts its knowledge into generalization. An opponent could still argue that it memorized first even if it lost the ability to output memorized images.
", "rawMarkdown": "Unfortunately there is no successful metric that distinguishes memorization from generalization which makes **proving** anything difficult.\n\nWhat if a GAN memorizes all the images first. And then it's put through a process that converts its knowledge into generalization. An opponent could still argue that it memorized first even if it lost the ability to output memorized images.", "votes": null }, { "id": "580192", "postDate": "07/19/2019 20:10:09", "content": "This kind of interesting community and environment. I've just joined.
", "rawMarkdown": "This kind of interesting community and environment. I've just joined.", "votes": null }, { "id": "580241", "postDate": "07/19/2019 21:59:58", "content": "I thought that . I am serious this quistion is good
", "rawMarkdown": "I thought that . I am serious this quistion is good", "votes": null }, { "id": "580364", "postDate": "07/20/2019 03:40:28", "content": "A very clear answer from Kaggle about the course of action they will take would be extremely important now. Myself I am holding back and using this competition just for experimentation with gans. Ideally they should check all the medal submissions, in my opinion.
", "rawMarkdown": "A very clear answer from Kaggle about the course of action they will take would be extremely important now. Myself I am holding back and using this competition just for experimentation with gans. Ideally they should check all the medal submissions, in my opinion.", "votes": null }, { "id": "580527", "postDate": "07/20/2019 09:54:05", "content": "@wendykan: Is it possible for you to provide insights regarding this matter? Thank you! :-)
", "rawMarkdown": "wendykan: Is it possible for you to provide insights regarding this matter? Thank you! :-)", "votes": null }, { "id": "580684", "postDate": "07/20/2019 15:08:08", "content": "as @cdeotte said, you can create a new GAN and justify it. \nWhat is a GAN? discriminator + generator? Many new GANs models were develop with the aim of improving others, minimize the FID metric in famous datasets such as celebA, cifar-10 ... why not do it here? memorizers are a great option. IMHO the problem is not our models or solutions, but the evaluation system they use.
\n\nPS. memorizer GANs are not perfect, they can choke in many situations ;)
", "rawMarkdown": "as @cdeotte said, you can create a new GAN and justify it. \n*What is a GAN? discriminator + generator?* Many new GANs models were develop with the aim of improving others, minimize the FID metric in famous datasets such as celebA, cifar-10 ... why not do it here? memorizers are a great option. IMHO the problem is not our models or solutions, but the evaluation system they use.\n\nPS. memorizer GANs are not perfect, they can choke in many situations ;)", "votes": null }, { "id": "580843", "postDate": "07/20/2019 21:43:30", "content": "You can create a new GAN, but if a GAN beats state of the art models in less than 9 hours of training, it is definitely fooling the metric. Best GANs are trained for days, the best researchers in the field have spent months developping them. We're competing on another level here.
\n\nIf anybody has found a paper considering the use of memorizers GAN for image generation, I'd love to read it. Because surely someone has thought about it already and realized : \"Oh look I can beat sota with this trick but it's useless !\"
", "rawMarkdown": "You can create a new GAN, but if a GAN beats state of the art models in less than 9 hours of training, it is definitely fooling the metric. Best GANs are trained for days, the best researchers in the field have spent months developping them. We're competing on another level here.\n\nIf anybody has found a paper considering the use of memorizers GAN for image generation, I'd love to read it. Because surely someone has thought about it already and realized : \"Oh look I can beat sota with this trick but it's useless !\"", "votes": null }, { "id": "580973", "postDate": "07/21/2019 07:22:13", "content": "Same here
", "rawMarkdown": "Same here", "votes": null }, { "id": "582893", "postDate": "07/23/2019 18:28:35", "content": "nada :/
", "rawMarkdown": "nada :/", "votes": null }, { "id": "582927", "postDate": "07/23/2019 19:40:56", "content": "totally agree.
", "rawMarkdown": "totally agree.", "votes": null }, { "id": "583321", "postDate": "07/24/2019 10:39:23", "content": "Nobody knows. I think the best thing you can do since you have 2 submissions is to select one with Memorizer GAN and another one with real GAN.
", "rawMarkdown": "Nobody knows. I think the best thing you can do since you have 2 submissions is to select one with Memorizer GAN and another one with real GAN.", "votes": null }, { "id": "583329", "postDate": "07/24/2019 10:44:24", "content": "I don't know if Kaggle can discard one of your submissions only though. It is likely that you can get removed totally from the leaderboard because one of your submissions is not legit, even if the other one is fine.
", "rawMarkdown": "I don't know if Kaggle can discard one of your submissions only though. It is likely that you can get removed totally from the leaderboard because one of your submissions is not legit, even if the other one is fine.", "votes": null }, { "id": "583352", "postDate": "07/24/2019 11:24:14", "content": "@theoviel Why? Not sure what could be the reason. Is it stated in the rules anywhere? (I know rules are the mess in this competition but still)
", "rawMarkdown": "theoviel Why? Not sure what could be the reason. Is it stated in the rules anywhere? (I know rules are the mess in this competition but still)", "votes": null }, { "id": "583358", "postDate": "07/24/2019 11:35:16", "content": "Why not? I'm just making assumptions here :)
", "rawMarkdown": "Why not? I'm just making assumptions here :)", "votes": null }, { "id": "583362", "postDate": "07/24/2019 11:43:19", "content": "May be you are right. Don't know.
", "rawMarkdown": "May be you are right. Don't know.", "votes": null }, { "id": "583367", "postDate": "07/24/2019 11:51:09", "content": "In between all legends of the game talking , I will give a coding newcomers view (3 months into python)\nIf we consider @dvorobiev is still using a \"Pure GAN \" ( He informed in another discussion thread that he is , when he was at 26 FiD) and his score is now 18-ish and is in 193rd position . Then based on what I have read so far in this competition , to achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .\nI have always been inspired by the novelty of 1st place , 2nd place solutions in Kaggle , for a change I will be delighted to read if someone publishes a \"My 300th place solution\" .
\n\nBeing said that , I am so far having a lot of fun, gaining a lot of knowledge as Its my 2nd competition ever . But at the end if I succumb to the extreme greed of using a memorizing GAN ( that tug-of-war in mind happens everyday, and I push it away everyday thinking , okay let me try for one more thing with vanilla GAN ) , I will only be blaming myself for taking that path . Not because I think Memorizer GANs are of lesser academic value , but because I would in that case probably will be resubmitting an available Kernel without adding any real value to myself or anyone.
", "rawMarkdown": "In between all legends of the game talking , I will give a coding newcomers view (3 months into python)\nIf we consider @dvorobiev is still using a \"Pure GAN \" ( He informed in another discussion thread that he is , when he was at 26 FiD) and his score is now 18-ish and is in 193rd position . Then based on what I have read so far in this competition , to achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .\nI have always been inspired by the novelty of 1st place , 2nd place solutions in Kaggle , for a change I will be delighted to read if someone publishes a \"My 300th place solution\" . \n\nBeing said that , I am so far having a lot of fun, gaining a lot of knowledge as Its my 2nd competition ever . But at the end if I succumb to the extreme greed of using a memorizing GAN ( that tug-of-war in mind happens everyday, and I push it away everyday thinking , okay let me try for one more thing with vanilla GAN ) , I will only be blaming myself for taking that path . Not because I think Memorizer GANs are of lesser academic value , but because I would in that case probably will be resubmitting an available Kernel without adding any real value to myself or anyone.", "votes": null }, { "id": "584884", "postDate": "07/26/2019 16:09:53", "content": "That's right.
\n\nCurrently GANs are slow and 9 hours limit is too little for real GAN. I am training GAN now outside this competition for another task and it takes several days each time.
\n\nMemorization GAN is just a trick. No model with N inputs and N outputs can have N parameters.
", "rawMarkdown": "That's right.\n\nCurrently GANs are slow and 9 hours limit is too little for real GAN. I am training GAN now outside this competition for another task and it takes several days each time. \n\nMemorization GAN is just a trick. No model with N inputs and N outputs can have N parameters.", "votes": null }, { "id": "585261", "postDate": "07/27/2019 07:13:50", "content": "\n\n\nto achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .
\n
I just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality.
", "rawMarkdown": ">to achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .\n\nI just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality.", "votes": null }, { "id": "585290", "postDate": "07/27/2019 07:55:34", "content": "\n\n\na score below 20
\n
Damn @dvorobiev ! is that really not a memorizing GAN?
", "rawMarkdown": ">a score below 20\n\nDamn @dvorobiev ! is that really not a memorizing GAN?", "votes": null }, { "id": "585292", "postDate": "07/27/2019 07:58:26", "content": "> I just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality.
\n\nThere is merit in showing up FID weakness as it has become the standard measure. I for one suspect a lot of new XYZGAN papers are doing the equivalent on CelebA as to us here with Dogs, getting 90, 60, 30 for images that don't look massively better. Beating the metric not making better gans.
\n\nHere is a paper that shows examples of better FID for subjectively worse images. \"To our best knowledge, we are the first to provide counter examples where FID gives inconsistent results with human judgments\".\nhttps://arxiv.org/abs/1803.07474\nhttps://github.com/B1ueber2y/CAFD
", "rawMarkdown": "> I just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality. \n\nThere is merit in showing up FID weakness as it has become the standard measure. I for one suspect a lot of new XYZGAN papers are doing the equivalent on CelebA as to us here with Dogs, getting 90, 60, 30 for images that don't look massively better. Beating the metric not making better gans.\n\nHere is a paper that shows examples of better FID for subjectively worse images. \"To our best knowledge, we are the first to provide counter examples where FID gives inconsistent results with human judgments\".\nhttps://arxiv.org/abs/1803.07474\nhttps://github.com/B1ueber2y/CAFD", "votes": null }, { "id": "585297", "postDate": "07/27/2019 08:16:08", "content": "@timetraveller98 \nIt's a conventional GAN :D \nBut there are more stuff than just a good model
", "rawMarkdown": "timetraveller98 \nIt's a conventional GAN :D \nBut there are more stuff than just a good model", "votes": null }, { "id": "585304", "postDate": "07/27/2019 08:32:39", "content": ".. and you are still in 225th with under 20 FiD :P .. lol ..
", "rawMarkdown": ".. and you are still in 225th with under 20 FiD :P .. lol ..", "votes": null }, { "id": "585487", "postDate": "07/27/2019 14:39:12", "content": "You can score bellow 30 using normal GANs so I would not play with Memorizers anymore, I will show this in a post/kernel soon and probably I will show how the mifid metric changes using another NN.\nI think you should keep trying and investigating about this amazing technology. \nPS: I'm on holidays, I wish they won't change the rules during the last week ;)
", "rawMarkdown": "You can score bellow 30 using normal GANs so I would not play with Memorizers anymore, I will show this in a post/kernel soon and probably I will show how the mifid metric changes using another NN.\nI think you should keep trying and investigating about this amazing technology. \nPS: I'm on holidays, I wish they won't change the rules during the last week ;)", "votes": null }, { "id": "585495", "postDate": "07/27/2019 14:48:04", "content": "Below 30 is not enough . Can it get to below 3 ? LoL
", "rawMarkdown": "Below 30 is not enough . Can it get to below 3 ? LoL", "votes": null }, { "id": "585678", "postDate": "07/27/2019 20:55:13", "content": "Great! @dvorobiev \nCan't wait to read your solution once the competition ends.
", "rawMarkdown": "Great! @dvorobiev \nCan't wait to read your solution once the competition ends.", "votes": null } ], "comments": [ { "id": 579911, "author_name": "cdeotte", "author_url": "", "post_date": "07/19/2019 11:45:50", "content": "Great question. There's also the fear of being incorrectly disqualified. If you invent a new kind of GAN that beats all published research paper GANs, it may get disqualified because the images look so good that Kaggle thinks you are memorizing training images (even if you are actually generalizing).
", "votes": null, "replies": [ { "id": 579915, "author_name": "abhishek", "author_url": "", "post_date": "07/19/2019 11:53:43", "content": "One can easily prove that and come back on the LB.
", "votes": null, "replies": [] }, { "id": 579926, "author_name": "cdeotte", "author_url": "", "post_date": "07/19/2019 12:03:43", "content": "Unfortunately there is no successful metric that distinguishes memorization from generalization which makes proving anything difficult.
\n\nWhat if a GAN memorizes all the images first. And then it's put through a process that converts its knowledge into generalization. An opponent could still argue that it memorized first even if it lost the ability to output memorized images.
", "votes": null, "replies": [] }, { "id": 583367, "author_name": "phoenix9032", "author_url": "", "post_date": "07/24/2019 11:51:09", "content": "In between all legends of the game talking , I will give a coding newcomers view (3 months into python)\nIf we consider @dvorobiev is still using a \"Pure GAN \" ( He informed in another discussion thread that he is , when he was at 26 FiD) and his score is now 18-ish and is in 193rd position . Then based on what I have read so far in this competition , to achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .\nI have always been inspired by the novelty of 1st place , 2nd place solutions in Kaggle , for a change I will be delighted to read if someone publishes a \"My 300th place solution\" .
\n\nBeing said that , I am so far having a lot of fun, gaining a lot of knowledge as Its my 2nd competition ever . But at the end if I succumb to the extreme greed of using a memorizing GAN ( that tug-of-war in mind happens everyday, and I push it away everyday thinking , okay let me try for one more thing with vanilla GAN ) , I will only be blaming myself for taking that path . Not because I think Memorizer GANs are of lesser academic value , but because I would in that case probably will be resubmitting an available Kernel without adding any real value to myself or anyone.
", "votes": null, "replies": [] }, { "id": 585261, "author_name": "dvorobiev", "author_url": "", "post_date": "07/27/2019 07:13:50", "content": "\n\n\nto achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .
\n
I just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality.
", "votes": null, "replies": [] }, { "id": 585290, "author_name": "timetraveller98", "author_url": "", "post_date": "07/27/2019 07:55:34", "content": "\n\n\na score below 20
\n
Damn @dvorobiev ! is that really not a memorizing GAN?
", "votes": null, "replies": [] }, { "id": 585292, "author_name": "robga", "author_url": "", "post_date": "07/27/2019 07:58:26", "content": "> I just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality.
\n\nThere is merit in showing up FID weakness as it has become the standard measure. I for one suspect a lot of new XYZGAN papers are doing the equivalent on CelebA as to us here with Dogs, getting 90, 60, 30 for images that don't look massively better. Beating the metric not making better gans.
\n\nHere is a paper that shows examples of better FID for subjectively worse images. \"To our best knowledge, we are the first to provide counter examples where FID gives inconsistent results with human judgments\".\nhttps://arxiv.org/abs/1803.07474\nhttps://github.com/B1ueber2y/CAFD
", "votes": null, "replies": [] }, { "id": 585297, "author_name": "dvorobiev", "author_url": "", "post_date": "07/27/2019 08:16:08", "content": "@timetraveller98 \nIt's a conventional GAN :D \nBut there are more stuff than just a good model
", "votes": null, "replies": [] }, { "id": 585304, "author_name": "phoenix9032", "author_url": "", "post_date": "07/27/2019 08:32:39", "content": ".. and you are still in 225th with under 20 FiD :P .. lol ..
", "votes": null, "replies": [] }, { "id": 585678, "author_name": "timetraveller98", "author_url": "", "post_date": "07/27/2019 20:55:13", "content": "Great! @dvorobiev \nCan't wait to read your solution once the competition ends.
", "votes": null, "replies": [] } ] }, { "id": 580192, "author_name": "dmsrus", "author_url": "", "post_date": "07/19/2019 20:10:09", "content": "This kind of interesting community and environment. I've just joined.
", "votes": null, "replies": [ { "id": 580973, "author_name": "heedlessstar852", "author_url": "", "post_date": "07/21/2019 07:22:13", "content": "Same here
", "votes": null, "replies": [] } ] }, { "id": 580241, "author_name": "batuhan35", "author_url": "", "post_date": "07/19/2019 21:59:58", "content": "I thought that . I am serious this quistion is good
", "votes": null, "replies": [] }, { "id": 580364, "author_name": "lucamassaron", "author_url": "", "post_date": "07/20/2019 03:40:28", "content": "A very clear answer from Kaggle about the course of action they will take would be extremely important now. Myself I am holding back and using this competition just for experimentation with gans. Ideally they should check all the medal submissions, in my opinion.
", "votes": null, "replies": [] }, { "id": 580527, "author_name": "projdev", "author_url": "", "post_date": "07/20/2019 09:54:05", "content": "@wendykan: Is it possible for you to provide insights regarding this matter? Thank you! :-)
", "votes": null, "replies": [] }, { "id": 580684, "author_name": "jesucristo", "author_url": "", "post_date": "07/20/2019 15:08:08", "content": "as @cdeotte said, you can create a new GAN and justify it. \nWhat is a GAN? discriminator + generator? Many new GANs models were develop with the aim of improving others, minimize the FID metric in famous datasets such as celebA, cifar-10 ... why not do it here? memorizers are a great option. IMHO the problem is not our models or solutions, but the evaluation system they use.
\n\nPS. memorizer GANs are not perfect, they can choke in many situations ;)
", "votes": null, "replies": [] }, { "id": 580843, "author_name": "theoviel", "author_url": "", "post_date": "07/20/2019 21:43:30", "content": "You can create a new GAN, but if a GAN beats state of the art models in less than 9 hours of training, it is definitely fooling the metric. Best GANs are trained for days, the best researchers in the field have spent months developping them. We're competing on another level here.
\n\nIf anybody has found a paper considering the use of memorizers GAN for image generation, I'd love to read it. Because surely someone has thought about it already and realized : \"Oh look I can beat sota with this trick but it's useless !\"
", "votes": null, "replies": [ { "id": 582927, "author_name": "zhangyixing0404", "author_url": "", "post_date": "07/23/2019 19:40:56", "content": "totally agree.
", "votes": null, "replies": [] }, { "id": 584884, "author_name": "demonplus", "author_url": "", "post_date": "07/26/2019 16:09:53", "content": "That's right.
\n\nCurrently GANs are slow and 9 hours limit is too little for real GAN. I am training GAN now outside this competition for another task and it takes several days each time.
\n\nMemorization GAN is just a trick. No model with N inputs and N outputs can have N parameters.
", "votes": null, "replies": [] } ] }, { "id": 582893, "author_name": "abhishek", "author_url": "", "post_date": "07/23/2019 18:28:35", "content": "nada :/
", "votes": null, "replies": [] }, { "id": 583321, "author_name": "demonplus", "author_url": "", "post_date": "07/24/2019 10:39:23", "content": "Nobody knows. I think the best thing you can do since you have 2 submissions is to select one with Memorizer GAN and another one with real GAN.
", "votes": null, "replies": [ { "id": 583329, "author_name": "theoviel", "author_url": "", "post_date": "07/24/2019 10:44:24", "content": "I don't know if Kaggle can discard one of your submissions only though. It is likely that you can get removed totally from the leaderboard because one of your submissions is not legit, even if the other one is fine.
", "votes": null, "replies": [] }, { "id": 583352, "author_name": "demonplus", "author_url": "", "post_date": "07/24/2019 11:24:14", "content": "@theoviel Why? Not sure what could be the reason. Is it stated in the rules anywhere? (I know rules are the mess in this competition but still)
", "votes": null, "replies": [] }, { "id": 583358, "author_name": "theoviel", "author_url": "", "post_date": "07/24/2019 11:35:16", "content": "Why not? I'm just making assumptions here :)
", "votes": null, "replies": [] }, { "id": 583362, "author_name": "demonplus", "author_url": "", "post_date": "07/24/2019 11:43:19", "content": "May be you are right. Don't know.
", "votes": null, "replies": [] } ] }, { "id": 585487, "author_name": "jesucristo", "author_url": "", "post_date": "07/27/2019 14:39:12", "content": "You can score bellow 30 using normal GANs so I would not play with Memorizers anymore, I will show this in a post/kernel soon and probably I will show how the mifid metric changes using another NN.\nI think you should keep trying and investigating about this amazing technology. \nPS: I'm on holidays, I wish they won't change the rules during the last week ;)
", "votes": null, "replies": [ { "id": 585495, "author_name": "phoenix9032", "author_url": "", "post_date": "07/27/2019 14:48:04", "content": "Below 30 is not enough . Can it get to below 3 ? LoL
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "579825": "I want to start working on this competition but before i invest my time in this one, I would really like to know what is Kaggle planning to do in the end.\n\n1. Will all the submissions not using a GAN approach be removed from LB?\n2. Will only the winner’s solution be verified?\n3. If the winners choose not to disclose their solution for forfeit prize money, will they be removed from LB?\n4. How many ranks will be checked for GAN based approach?\n\nI really don’t want to end up wasting my time because of a bunch of people who didn’t follow the rules.", "579911": "Great question. There's also the fear of being incorrectly disqualified. If you invent a new kind of GAN that beats all published research paper GANs, it may get disqualified because the images look so good that Kaggle thinks you are memorizing training images (even if you are actually generalizing).", "579915": "One can easily prove that and come back on the LB.", "579926": "Unfortunately there is no successful metric that distinguishes memorization from generalization which makes **proving** anything difficult.\n\nWhat if a GAN memorizes all the images first. And then it's put through a process that converts its knowledge into generalization. An opponent could still argue that it memorized first even if it lost the ability to output memorized images.", "580192": "This kind of interesting community and environment. I've just joined.", "580241": "I thought that . I am serious this quistion is good", "580364": "A very clear answer from Kaggle about the course of action they will take would be extremely important now. Myself I am holding back and using this competition just for experimentation with gans. Ideally they should check all the medal submissions, in my opinion.", "580527": "wendykan: Is it possible for you to provide insights regarding this matter? Thank you! :-)", "580684": "as @cdeotte said, you can create a new GAN and justify it. \n*What is a GAN? discriminator + generator?* Many new GANs models were develop with the aim of improving others, minimize the FID metric in famous datasets such as celebA, cifar-10 ... why not do it here? memorizers are a great option. IMHO the problem is not our models or solutions, but the evaluation system they use.\n\nPS. memorizer GANs are not perfect, they can choke in many situations ;)", "580843": "You can create a new GAN, but if a GAN beats state of the art models in less than 9 hours of training, it is definitely fooling the metric. Best GANs are trained for days, the best researchers in the field have spent months developping them. We're competing on another level here.\n\nIf anybody has found a paper considering the use of memorizers GAN for image generation, I'd love to read it. Because surely someone has thought about it already and realized : \"Oh look I can beat sota with this trick but it's useless !\"", "580973": "Same here", "582893": "nada :/", "582927": "totally agree.", "583321": "Nobody knows. I think the best thing you can do since you have 2 submissions is to select one with Memorizer GAN and another one with real GAN.", "583329": "I don't know if Kaggle can discard one of your submissions only though. It is likely that you can get removed totally from the leaderboard because one of your submissions is not legit, even if the other one is fine.", "583352": "theoviel Why? Not sure what could be the reason. Is it stated in the rules anywhere? (I know rules are the mess in this competition but still)", "583358": "Why not? I'm just making assumptions here :)", "583362": "May be you are right. Don't know.", "583367": "In between all legends of the game talking , I will give a coding newcomers view (3 months into python)\nIf we consider @dvorobiev is still using a \"Pure GAN \" ( He informed in another discussion thread that he is , when he was at 26 FiD) and his score is now 18-ish and is in 193rd position . Then based on what I have read so far in this competition , to achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .\nI have always been inspired by the novelty of 1st place , 2nd place solutions in Kaggle , for a change I will be delighted to read if someone publishes a \"My 300th place solution\" . \n\nBeing said that , I am so far having a lot of fun, gaining a lot of knowledge as Its my 2nd competition ever . But at the end if I succumb to the extreme greed of using a memorizing GAN ( that tug-of-war in mind happens everyday, and I push it away everyday thinking , okay let me try for one more thing with vanilla GAN ) , I will only be blaming myself for taking that path . Not because I think Memorizer GANs are of lesser academic value , but because I would in that case probably will be resubmitting an available Kernel without adding any real value to myself or anyone.", "584884": "That's right.\n\nCurrently GANs are slow and 9 hours limit is too little for real GAN. I am training GAN now outside this competition for another task and it takes several days each time. \n\nMemorization GAN is just a trick. No model with N inputs and N outputs can have N parameters.", "585261": ">to achieve that in 9 hours calls for a new academic material itself, which I am very very eager to read .\n\nI just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality.", "585290": ">a score below 20\n\nDamn @dvorobiev ! is that really not a memorizing GAN?", "585292": "> I just tweaked some parameters to get a score below 20. But the images looks mostly the same. There are still plenty of crappy ones. I guess, I am optimizing my model for metric itself rather than for real quality. \n\nThere is merit in showing up FID weakness as it has become the standard measure. I for one suspect a lot of new XYZGAN papers are doing the equivalent on CelebA as to us here with Dogs, getting 90, 60, 30 for images that don't look massively better. Beating the metric not making better gans.\n\nHere is a paper that shows examples of better FID for subjectively worse images. \"To our best knowledge, we are the first to provide counter examples where FID gives inconsistent results with human judgments\".\nhttps://arxiv.org/abs/1803.07474\nhttps://github.com/B1ueber2y/CAFD", "585297": "timetraveller98 \nIt's a conventional GAN :D \nBut there are more stuff than just a good model", "585304": ".. and you are still in 225th with under 20 FiD :P .. lol ..", "585487": "You can score bellow 30 using normal GANs so I would not play with Memorizers anymore, I will show this in a post/kernel soon and probably I will show how the mifid metric changes using another NN.\nI think you should keep trying and investigating about this amazing technology. \nPS: I'm on holidays, I wish they won't change the rules during the last week ;)", "585495": "Below 30 is not enough . Can it get to below 3 ? LoL", "585678": "Great! @dvorobiev \nCan't wait to read your solution once the competition ends." }, "source": "meta" }