Hi Kagglers ,\nI have made my Kernel below as Public now . It scores 67 FiD in LB . I have tweaked it a little \nwith different kind of normalization concept from Progressive GANs . \nIf anyone new like me is interested to refer ,fork ,run and learn . Please go ahead . Also please let me know \nyour feedback if you get a better score .\nPlease upvote if it helps your learning and give you more motivation to stick in this competition.\nMy Kernel with 67 MfId
\n\nRegards,\nNirjhar
\n\nUpdate : The version 3 scores 62 now . No other changes . Just ran for longer . \nUpdate 2: Latest version with 2000 epochs goes back to 67 . So the optimum is around 1400 with this setup .
", "messages": [ { "id": "584126", "postDate": "07/25/2019 13:18:07", "content": "Hi Kagglers ,\nI have made my Kernel below as Public now . It scores 67 FiD in LB . I have tweaked it a little \nwith different kind of normalization concept from Progressive GANs . \nIf anyone new like me is interested to refer ,fork ,run and learn . Please go ahead . Also please let me know \nyour feedback if you get a better score .\nPlease upvote if it helps your learning and give you more motivation to stick in this competition.\nMy Kernel with 67 MfId
\n\nRegards,\nNirjhar
\n\nUpdate : The version 3 scores 62 now . No other changes . Just ran for longer . \nUpdate 2: Latest version with 2000 epochs goes back to 67 . So the optimum is around 1400 with this setup .
", "rawMarkdown": "Hi Kagglers ,\nI have made my Kernel below as Public now . It scores 67 FiD in LB . I have tweaked it a little \nwith different kind of normalization concept from Progressive GANs . \nIf anyone new like me is interested to refer ,fork ,run and learn . Please go ahead . Also please let me know \nyour feedback if you get a better score .\nPlease upvote if it helps your learning and give you more motivation to stick in this competition.\n[My Kernel with 67 MfId](https://www.kaggle.com/phoenix9032/gan-dogs-starter-24-jul-custom-layers)\n\nRegards,\nNirjhar \n\nUpdate : The version 3 scores 62 now . No other changes . Just ran for longer . \nUpdate 2: Latest version with 2000 epochs goes back to 67 . So the optimum is around 1400 with this setup .", "votes": null }, { "id": "584233", "postDate": "07/25/2019 15:46:51", "content": "Thanks for your kernel!\nHave u tried adding conditions?
", "rawMarkdown": "Thanks for your kernel!\nHave u tried adding conditions?", "votes": null }, { "id": "584259", "postDate": "07/25/2019 16:18:44", "content": "Great! I've observed truncated norm improves the score by 5-6 points too. My current best real GAN score is 71, still figuring things out. Your kernel will be helpful.
", "rawMarkdown": "Great! I've observed truncated norm improves the score by 5-6 points too. My current best real GAN score is 71, still figuring things out. Your kernel will be helpful.", "votes": null }, { "id": "584346", "postDate": "07/25/2019 19:21:52", "content": "Thank you .. If you are able to find better score by tweaking it .. do drop a hint . Let me know if there are any issues too .
", "rawMarkdown": "Thank you .. If you are able to find better score by tweaking it .. do drop a hint . Let me know if there are any issues too .", "votes": null }, { "id": "584347", "postDate": "07/25/2019 19:22:25", "content": "I am really sorry , I could not understand you . Are you talking about Conditional GAN?
", "rawMarkdown": "I am really sorry , I could not understand you . Are you talking about Conditional GAN?", "votes": null }, { "id": "584367", "postDate": "07/25/2019 20:33:46", "content": "Congrats! I think it's a very good score for a handcrafted architecture
", "rawMarkdown": "Congrats! I think it's a very good score for a handcrafted architecture", "votes": null }, { "id": "584368", "postDate": "07/25/2019 20:35:57", "content": "I tried it, and it did improve my results for not-fully converged net. But for a long run (about 8 hours) the gain was negligible (~ 0.5)
", "rawMarkdown": "I tried it, and it did improve my results for not-fully converged net. But for a long run (about 8 hours) the gain was negligible (~ 0.5)", "votes": null }, { "id": "584371", "postDate": "07/25/2019 20:41:43", "content": "This made my day !! Thank you .
", "rawMarkdown": "This made my day !! Thank you .", "votes": null }, { "id": "584378", "postDate": "07/25/2019 20:59:43", "content": "yep, adding vector of breed in addition to the image
", "rawMarkdown": "yep, adding vector of breed in addition to the image", "votes": null }, { "id": "584382", "postDate": "07/25/2019 21:02:20", "content": "Oh . I have not tried CGAN yet . I am still working on vanilla GANs . But , wanted to give it a try before competition ends .
", "rawMarkdown": "Oh . I have not tried CGAN yet . I am still working on vanilla GANs . But , wanted to give it a try before competition ends .", "votes": null }, { "id": "584594", "postDate": "07/26/2019 07:49:48", "content": "Great work @phoenix9032 and congrats on your score.
", "rawMarkdown": "Great work @phoenix9032 and congrats on your score.", "votes": null }, { "id": "585076", "postDate": "07/26/2019 22:31:05", "content": "Thank you. But surprisingly when I tried the new normalization and other new layers, the score worsened a little bit. Maybe I did something wrong. ☹️
", "rawMarkdown": "Thank you. But surprisingly when I tried the new normalization and other new layers, the score worsened a little bit. Maybe I did something wrong. ☹️", "votes": null }, { "id": "585148", "postDate": "07/27/2019 03:05:44", "content": "Great job Nirjhar. LB 62 is a fantastic job.
", "rawMarkdown": "Great job Nirjhar. LB 62 is a fantastic job.", "votes": null }, { "id": "585534", "postDate": "07/27/2019 15:50:23", "content": "Dear Chris ,
\n\nWith a little more tweak of this I broke into 57 range .
\n\nRegards,\nNirjhar
", "rawMarkdown": "Dear Chris ,\n\nWith a little more tweak of this I broke into 57 range . \n\nRegards,\nNirjhar", "votes": null }, { "id": "585818", "postDate": "07/28/2019 06:16:55", "content": "btw, i`ve tried it, score got little bit worse
", "rawMarkdown": "btw, i`ve tried it, score got little bit worse", "votes": null }, { "id": "586440", "postDate": "07/29/2019 07:13:11", "content": "its surprising that additional class information doesn't help that much :(
", "rawMarkdown": "its surprising that additional class information doesn't help that much :(", "votes": null }, { "id": "586689", "postDate": "07/29/2019 14:31:28", "content": "May be because FiD penalizes the images too similar to training images , and by generating classwise sample are we reducing the diversity somehow ?
", "rawMarkdown": "May be because FiD penalizes the images too similar to training images , and by generating classwise sample are we reducing the diversity somehow ?", "votes": null }, { "id": "588880", "postDate": "07/31/2019 07:17:45", "content": "Great kernel, really helped me a lot, I used this as a baseline for my solution. Hopefully, soon I will upload a version of my solution, which is essentially a more organized version of your solution.
", "rawMarkdown": "Great kernel, really helped me a lot, I used this as a baseline for my solution. Hopefully, soon I will upload a version of my solution, which is essentially a more organized version of your solution.", "votes": null }, { "id": "588935", "postDate": "07/31/2019 08:54:35", "content": "Thanks a lot . Glad to know . Looking forward to it Shay.
", "rawMarkdown": "Thanks a lot . Glad to know . Looking forward to it Shay.", "votes": null } ], "comments": [ { "id": 584233, "author_name": "izmaylov", "author_url": "", "post_date": "07/25/2019 15:46:51", "content": "Thanks for your kernel!\nHave u tried adding conditions?
", "votes": null, "replies": [ { "id": 584347, "author_name": "phoenix9032", "author_url": "", "post_date": "07/25/2019 19:22:25", "content": "I am really sorry , I could not understand you . Are you talking about Conditional GAN?
", "votes": null, "replies": [] }, { "id": 584378, "author_name": "izmaylov", "author_url": "", "post_date": "07/25/2019 20:59:43", "content": "yep, adding vector of breed in addition to the image
", "votes": null, "replies": [] }, { "id": 584382, "author_name": "phoenix9032", "author_url": "", "post_date": "07/25/2019 21:02:20", "content": "Oh . I have not tried CGAN yet . I am still working on vanilla GANs . But , wanted to give it a try before competition ends .
", "votes": null, "replies": [] }, { "id": 585818, "author_name": "izmaylov", "author_url": "", "post_date": "07/28/2019 06:16:55", "content": "btw, i`ve tried it, score got little bit worse
", "votes": null, "replies": [] }, { "id": 586440, "author_name": "yangsaewon", "author_url": "", "post_date": "07/29/2019 07:13:11", "content": "its surprising that additional class information doesn't help that much :(
", "votes": null, "replies": [] }, { "id": 586689, "author_name": "phoenix9032", "author_url": "", "post_date": "07/29/2019 14:31:28", "content": "May be because FiD penalizes the images too similar to training images , and by generating classwise sample are we reducing the diversity somehow ?
", "votes": null, "replies": [] } ] }, { "id": 584259, "author_name": "timetraveller98", "author_url": "", "post_date": "07/25/2019 16:18:44", "content": "Great! I've observed truncated norm improves the score by 5-6 points too. My current best real GAN score is 71, still figuring things out. Your kernel will be helpful.
", "votes": null, "replies": [ { "id": 584346, "author_name": "phoenix9032", "author_url": "", "post_date": "07/25/2019 19:21:52", "content": "Thank you .. If you are able to find better score by tweaking it .. do drop a hint . Let me know if there are any issues too .
", "votes": null, "replies": [] }, { "id": 584368, "author_name": "dvorobiev", "author_url": "", "post_date": "07/25/2019 20:35:57", "content": "I tried it, and it did improve my results for not-fully converged net. But for a long run (about 8 hours) the gain was negligible (~ 0.5)
", "votes": null, "replies": [] } ] }, { "id": 584367, "author_name": "dvorobiev", "author_url": "", "post_date": "07/25/2019 20:33:46", "content": "Congrats! I think it's a very good score for a handcrafted architecture
", "votes": null, "replies": [ { "id": 584371, "author_name": "phoenix9032", "author_url": "", "post_date": "07/25/2019 20:41:43", "content": "This made my day !! Thank you .
", "votes": null, "replies": [] } ] }, { "id": 584594, "author_name": "joelhanson", "author_url": "", "post_date": "07/26/2019 07:49:48", "content": "Great work @phoenix9032 and congrats on your score.
", "votes": null, "replies": [] }, { "id": 585076, "author_name": "ajinomoto132", "author_url": "", "post_date": "07/26/2019 22:31:05", "content": "Thank you. But surprisingly when I tried the new normalization and other new layers, the score worsened a little bit. Maybe I did something wrong. ☹️
", "votes": null, "replies": [] }, { "id": 585148, "author_name": "cdeotte", "author_url": "", "post_date": "07/27/2019 03:05:44", "content": "Great job Nirjhar. LB 62 is a fantastic job.
", "votes": null, "replies": [ { "id": 585534, "author_name": "phoenix9032", "author_url": "", "post_date": "07/27/2019 15:50:23", "content": "Dear Chris ,
\n\nWith a little more tweak of this I broke into 57 range .
\n\nRegards,\nNirjhar
", "votes": null, "replies": [] } ] }, { "id": 588880, "author_name": "shaygu", "author_url": "", "post_date": "07/31/2019 07:17:45", "content": "Great kernel, really helped me a lot, I used this as a baseline for my solution. Hopefully, soon I will upload a version of my solution, which is essentially a more organized version of your solution.
", "votes": null, "replies": [ { "id": 588935, "author_name": "phoenix9032", "author_url": "", "post_date": "07/31/2019 08:54:35", "content": "Thanks a lot . Glad to know . Looking forward to it Shay.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "584126": "Hi Kagglers ,\nI have made my Kernel below as Public now . It scores 67 FiD in LB . I have tweaked it a little \nwith different kind of normalization concept from Progressive GANs . \nIf anyone new like me is interested to refer ,fork ,run and learn . Please go ahead . Also please let me know \nyour feedback if you get a better score .\nPlease upvote if it helps your learning and give you more motivation to stick in this competition.\n[My Kernel with 67 MfId](https://www.kaggle.com/phoenix9032/gan-dogs-starter-24-jul-custom-layers)\n\nRegards,\nNirjhar \n\nUpdate : The version 3 scores 62 now . No other changes . Just ran for longer . \nUpdate 2: Latest version with 2000 epochs goes back to 67 . So the optimum is around 1400 with this setup .", "584233": "Thanks for your kernel!\nHave u tried adding conditions?", "584259": "Great! I've observed truncated norm improves the score by 5-6 points too. My current best real GAN score is 71, still figuring things out. Your kernel will be helpful.", "584346": "Thank you .. If you are able to find better score by tweaking it .. do drop a hint . Let me know if there are any issues too .", "584347": "I am really sorry , I could not understand you . Are you talking about Conditional GAN?", "584367": "Congrats! I think it's a very good score for a handcrafted architecture", "584368": "I tried it, and it did improve my results for not-fully converged net. But for a long run (about 8 hours) the gain was negligible (~ 0.5)", "584371": "This made my day !! Thank you .", "584378": "yep, adding vector of breed in addition to the image", "584382": "Oh . I have not tried CGAN yet . I am still working on vanilla GANs . But , wanted to give it a try before competition ends .", "584594": "Great work @phoenix9032 and congrats on your score.", "585076": "Thank you. But surprisingly when I tried the new normalization and other new layers, the score worsened a little bit. Maybe I did something wrong. ☹️", "585148": "Great job Nirjhar. LB 62 is a fantastic job.", "585534": "Dear Chris ,\n\nWith a little more tweak of this I broke into 57 range . \n\nRegards,\nNirjhar", "585818": "btw, i`ve tried it, score got little bit worse", "586440": "its surprising that additional class information doesn't help that much :(", "586689": "May be because FiD penalizes the images too similar to training images , and by generating classwise sample are we reducing the diversity somehow ?", "588880": "Great kernel, really helped me a lot, I used this as a baseline for my solution. Hopefully, soon I will upload a version of my solution, which is essentially a more organized version of your solution.", "588935": "Thanks a lot . Glad to know . Looking forward to it Shay." }, "source": "meta" }