For those who advanced 500+ place on the private leaderboard, what did you do?
", "messages": [ { "id": "676054", "postDate": "11/19/2019 00:06:18", "content": "For those who advanced 500+ place on the private leaderboard, what did you do?
", "rawMarkdown": "For those who advanced 500+ place on the private leaderboard, what did you do?", "votes": null }, { "id": "676062", "postDate": "11/19/2019 00:15:09", "content": "I dropped 605 places :( My mistake was I reduced the holdout size too much and overfit the Public leaderboard, my best private results would have kept me at around 300th place
", "rawMarkdown": "I dropped 605 places :( My mistake was I reduced the holdout size too much and overfit the Public leaderboard, my best private results would have kept me at around 300th place", "votes": null }, { "id": "676071", "postDate": "11/19/2019 00:23:32", "content": "yikes!
", "rawMarkdown": "yikes!", "votes": null }, { "id": "676085", "postDate": "11/19/2019 00:32:35", "content": "They are The Flash (superhuman speeds).
", "rawMarkdown": "They are The Flash (superhuman speeds).", "votes": null }, { "id": "676089", "postDate": "11/19/2019 00:34:19", "content": "i think here is same. We used 11 folds :) 1 as hold out.
", "rawMarkdown": "i think here is same. We used 11 folds :) 1 as hold out.", "votes": null }, { "id": "676218", "postDate": "11/19/2019 03:08:14", "content": "I think you are referring me XD\nActually I just ensemble six models (each run with 3-fold), using another GCN multilabel classifier to remove false masks.
", "rawMarkdown": "I think you are referring me XD\nActually I just ensemble six models (each run with 3-fold), using another GCN multilabel classifier to remove false masks.", "votes": null }, { "id": "676410", "postDate": "11/19/2019 06:34:42", "content": "I had pretty common approach nothing fancy, got 0.6565 on public LB using efficientnet-b4 (BCEDice Loss) and normal post processing technique. Ensemble was never able to cross that score nor any new approaches were able to. I didn't change threshold or min area much so that I don't overfit public LB that is it. And got Private LB 0.65477 and +533. I never expected this to happen.
", "rawMarkdown": "I had pretty common approach nothing fancy, got 0.6565 on public LB using efficientnet-b4 (BCEDice Loss) and normal post processing technique. Ensemble was never able to cross that score nor any new approaches were able to. I didn't change threshold or min area much so that I don't overfit public LB that is it. And got Private LB 0.65477 and +533. I never expected this to happen.", "votes": null }, { "id": "677079", "postDate": "11/19/2019 19:01:07", "content": "@joeychuang By ensamble six models (each run with 3-fold) do you mean that you make 6 different types of architectures(u-net vs fpn, resnet50 vs rse_resnext101_32x4d, batch size..etc), then in each one you split the data in 3 folds (train on 1 and 2, test on 3....train on 2 and 3,test on 1..... train on 1 and 3, test on 2), then make the mean of the 3 folds results for each of the 9 clasifiers and then make the mean of the 9 clasifiers?
\n\nThanks
", "rawMarkdown": "joeychuang By ensamble six models (each run with 3-fold) do you mean that you make 6 different types of architectures(u-net vs fpn, resnet50 vs rse_resnext101_32x4d, batch size..etc), then in each one you split the data in 3 folds (train on 1 and 2, test on 3....train on 2 and 3,test on 1..... train on 1 and 3, test on 2), then make the mean of the 3 folds results for each of the 9 clasifiers and then make the mean of the 9 clasifiers? \n\nThanks", "votes": null }, { "id": "677322", "postDate": "11/20/2019 03:32:32", "content": "Almost like this! I also tried different augmentation.
", "rawMarkdown": "Almost like this! I also tried different augmentation.", "votes": null }, { "id": "677557", "postDate": "11/20/2019 10:16:46", "content": "Nice ! Congratulations for you place !
", "rawMarkdown": "Nice ! Congratulations for you place !", "votes": null } ], "comments": [ { "id": 676062, "author_name": "s13658695", "author_url": "", "post_date": "11/19/2019 00:15:09", "content": "I dropped 605 places :( My mistake was I reduced the holdout size too much and overfit the Public leaderboard, my best private results would have kept me at around 300th place
", "votes": null, "replies": [ { "id": 676071, "author_name": "ekan825", "author_url": "", "post_date": "11/19/2019 00:23:32", "content": "yikes!
", "votes": null, "replies": [] }, { "id": 676089, "author_name": "tugstugi", "author_url": "", "post_date": "11/19/2019 00:34:19", "content": "i think here is same. We used 11 folds :) 1 as hold out.
", "votes": null, "replies": [] } ] }, { "id": 676085, "author_name": "mpwolke", "author_url": "", "post_date": "11/19/2019 00:32:35", "content": "They are The Flash (superhuman speeds).
", "votes": null, "replies": [] }, { "id": 676218, "author_name": "joeychuang", "author_url": "", "post_date": "11/19/2019 03:08:14", "content": "I think you are referring me XD\nActually I just ensemble six models (each run with 3-fold), using another GCN multilabel classifier to remove false masks.
", "votes": null, "replies": [] }, { "id": 676410, "author_name": "hansol1sol", "author_url": "", "post_date": "11/19/2019 06:34:42", "content": "I had pretty common approach nothing fancy, got 0.6565 on public LB using efficientnet-b4 (BCEDice Loss) and normal post processing technique. Ensemble was never able to cross that score nor any new approaches were able to. I didn't change threshold or min area much so that I don't overfit public LB that is it. And got Private LB 0.65477 and +533. I never expected this to happen.
", "votes": null, "replies": [] }, { "id": 677079, "author_name": "vladvdv", "author_url": "", "post_date": "11/19/2019 19:01:07", "content": "@joeychuang By ensamble six models (each run with 3-fold) do you mean that you make 6 different types of architectures(u-net vs fpn, resnet50 vs rse_resnext101_32x4d, batch size..etc), then in each one you split the data in 3 folds (train on 1 and 2, test on 3....train on 2 and 3,test on 1..... train on 1 and 3, test on 2), then make the mean of the 3 folds results for each of the 9 clasifiers and then make the mean of the 9 clasifiers?
\n\nThanks
", "votes": null, "replies": [ { "id": 677322, "author_name": "joeychuang", "author_url": "", "post_date": "11/20/2019 03:32:32", "content": "Almost like this! I also tried different augmentation.
", "votes": null, "replies": [] }, { "id": 677557, "author_name": "vladvdv", "author_url": "", "post_date": "11/20/2019 10:16:46", "content": "Nice ! Congratulations for you place !
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "676054": "For those who advanced 500+ place on the private leaderboard, what did you do?", "676062": "I dropped 605 places :( My mistake was I reduced the holdout size too much and overfit the Public leaderboard, my best private results would have kept me at around 300th place", "676071": "yikes!", "676085": "They are The Flash (superhuman speeds).", "676089": "i think here is same. We used 11 folds :) 1 as hold out.", "676218": "I think you are referring me XD\nActually I just ensemble six models (each run with 3-fold), using another GCN multilabel classifier to remove false masks.", "676410": "I had pretty common approach nothing fancy, got 0.6565 on public LB using efficientnet-b4 (BCEDice Loss) and normal post processing technique. Ensemble was never able to cross that score nor any new approaches were able to. I didn't change threshold or min area much so that I don't overfit public LB that is it. And got Private LB 0.65477 and +533. I never expected this to happen.", "677079": "joeychuang By ensamble six models (each run with 3-fold) do you mean that you make 6 different types of architectures(u-net vs fpn, resnet50 vs rse_resnext101_32x4d, batch size..etc), then in each one you split the data in 3 folds (train on 1 and 2, test on 3....train on 2 and 3,test on 1..... train on 1 and 3, test on 2), then make the mean of the 3 folds results for each of the 9 clasifiers and then make the mean of the 9 clasifiers? \n\nThanks", "677322": "Almost like this! I also tried different augmentation.", "677557": "Nice ! Congratulations for you place !" }, "source": "meta" }