![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1341302%2F9b0a79be9acf60ceeadbe6fc41451a32%2Fbest_score.PNG?generation=1574124648184422&alt=media\")
Just curious...
", "messages": [ { "id": "676067", "postDate": "11/19/2019 00:19:16", "content": "Just curious...
", "rawMarkdown": "Just curious...", "votes": null }, { "id": "676069", "postDate": "11/19/2019 00:22:36", "content": "0.66154.. very close to my final submission. Big ensemble save me from the shake up Lol.\nBut shake up seems quite small in this competition.
", "rawMarkdown": "0.66154.. very close to my final submission. Big ensemble save me from the shake up Lol.\nBut shake up seems quite small in this competition.", "votes": null }, { "id": "676072", "postDate": "11/19/2019 00:23:42", "content": "only 0.66104. fall down 115 places :) last day started to tune threshold...
", "rawMarkdown": "only 0.66104. fall down 115 places :) last day started to tune threshold...", "votes": null }, { "id": "676073", "postDate": "11/19/2019 00:25:51", "content": "0.65725. I used a different classifier
", "rawMarkdown": "0.65725. I used a different classifier", "votes": null }, { "id": "676077", "postDate": "11/19/2019 00:26:52", "content": "private LB 0.66334, public LB 0.66883
", "rawMarkdown": "private LB 0.66334, public LB 0.66883", "votes": null }, { "id": "676079", "postDate": "11/19/2019 00:28:16", "content": "0.66824 (would be 9th place I think)\nthe interesting part is: this solution was an emsemble with models trained for each class that I thought was not working well
", "rawMarkdown": "0.66824 (would be 9th place I think)\nthe interesting part is: this solution was an emsemble with models trained for each class that I thought was not working well", "votes": null }, { "id": "676083", "postDate": "11/19/2019 00:31:20", "content": "My best submission was the one I picked. 20 small model (resnet18, resnet 34, efficientnetb0) ensemble with fixed max pixel threshold and fixed 0.3 mask threshold.
", "rawMarkdown": "My best submission was the one I picked. 20 small model (resnet18, resnet 34, efficientnetb0) ensemble with fixed max pixel threshold and fixed 0.3 mask threshold.", "votes": null }, { "id": "676103", "postDate": "11/19/2019 00:50:49", "content": "
tough one to digest. I'm sorry for your lost gold. Hope it will be a nice learning experience for you for future competitions
", "rawMarkdown": "tough one to digest. I'm sorry for your lost gold. Hope it will be a nice learning experience for you for future competitions", "votes": null }, { "id": "676135", "postDate": "11/19/2019 01:38:35", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1175497%2F5c0938faabae9ab82b673abc0a0da3d1%2FScreenshot%20from%202019-11-18%2020-36-05.png?generation=1574127399343412&alt=media\")
Just tried one more submission after the competition ended.
", "rawMarkdown": "\n\nJust tried one more submission after the competition ended.", "votes": null }, { "id": "676136", "postDate": "11/19/2019 01:39:23", "content": "I learned my lesson from Steel comp and chose correctly this comp. I purposely chose less than my best public LB. I chose my best CV of 0.663 (with Public LB 0.670) which has Private LB 0.663. My best Public LB of 0.672 has Private LB 0.6615.
", "rawMarkdown": "I learned my lesson from Steel comp and chose correctly this comp. I purposely chose less than my best public LB. I chose my best CV of 0.663 (with Public LB 0.670) which has Private LB 0.663. My best Public LB of 0.672 has Private LB 0.6615.", "votes": null }, { "id": "676144", "postDate": "11/19/2019 01:47:01", "content": "This competition , even though noisy data , the CV and LB correlated quite well ... So choosing based on CV was the right thing to do :) .. and no major shake up :D
", "rawMarkdown": "This competition , even though noisy data , the CV and LB correlated quite well ... So choosing based on CV was the right thing to do :) .. and no major shake up :D", "votes": null }, { "id": "676149", "postDate": "11/19/2019 01:49:15", "content": "Yes. Congratulations Nijhar and team. You all did awesome!! What models did you use? I saw that you jumped many positions in the last few days.
", "rawMarkdown": "Yes. Congratulations Nijhar and team. You all did awesome!! What models did you use? I saw that you jumped many positions in the last few days.", "votes": null }, { "id": "676157", "postDate": "11/19/2019 01:59:31", "content": "This surprised me, bro. This is like the opposite of my situation...\nBut any reason about your determination?
", "rawMarkdown": "This surprised me, bro. This is like the opposite of my situation...\nBut any reason about your determination?", "votes": null }, { "id": "676168", "postDate": "11/19/2019 02:12:25", "content": "OK this is me. At least we know up to 3 digits difference is untrustworthy.
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F274731%2Fa1de7b099935bd6032b59f185f33d736%2F1574123223672.jpg?generation=1574128799973828&alt=media\")
And I always used threshold=0.6 in the ensemble, that's about 38% positive, the only submission with lower threshold=0.55 gave best score. (I estimate it to be about 40% positive)
", "rawMarkdown": "And I always used threshold=0.6 in the ensemble, that's about 38% positive, the only submission with lower threshold=0.55 gave best score. (I estimate it to be about 40% positive)", "votes": null }, { "id": "676173", "postDate": "11/19/2019 02:17:57", "content": "I had a feeling that my Public LB 0.672 model was overfitting. When my LB position was at 0.670, I did a few weird tricks and my score increased to LB 0.671. Then more tricks increased it to LB 0.672. But those same tricks didn't increase my CV, so I was suspicious that model was overfitting LB.
\n\nMy other model always had CV and LB change together.
", "rawMarkdown": "I had a feeling that my Public LB 0.672 model was overfitting. When my LB position was at 0.670, I did a few weird tricks and my score increased to LB 0.671. Then more tricks increased it to LB 0.672. But those same tricks didn't increase my CV, so I was suspicious that model was overfitting LB.\n\nMy other model always had CV and LB change together.", "votes": null }, { "id": "676177", "postDate": "11/19/2019 02:21:04", "content": "Oh no, bad thing happens sometimes. And good luck next competition!
", "rawMarkdown": "Oh no, bad thing happens sometimes. And good luck next competition!", "votes": null }, { "id": "676180", "postDate": "11/19/2019 02:24:01", "content": "Lovazloss didn't work for me this time, and I hardly had success with inception, looking forward to your write up!
", "rawMarkdown": "Lovazloss didn't work for me this time, and I hardly had success with inception, looking forward to your write up!", "votes": null }, { "id": "676181", "postDate": "11/19/2019 02:25:53", "content": "Thanks. Trust your CV, lessons learned.
", "rawMarkdown": "Thanks. Trust your CV, lessons learned.", "votes": null }, { "id": "676208", "postDate": "11/19/2019 02:55:04", "content": "None of us actually thought it's possible to jump . But all credit goes to @kani23 . He was able to make Heng's model work and improve in last 2 days and we got a surprise improvement . When he gets up , he can explain first hand :)
", "rawMarkdown": "None of us actually thought it's possible to jump . But all credit goes to @kani23 . He was able to make Heng's model work and improve in last 2 days and we got a surprise improvement . When he gets up , he can explain first hand :)", "votes": null }, { "id": "676212", "postDate": "11/19/2019 03:01:10", "content": "\n\n\nThanks. Trust your CV, lessons learned.
\n
Not always in Kaggle comps. You need to determine where the test data comes from. In this comp, the research paper indicates that there is one big dataset and they split it between train and test. But in other comps, sometimes they use different datasets for test and train, and then you need to probe and watch LB more.
", "rawMarkdown": "> Thanks. Trust your CV, lessons learned.\n\nNot always in Kaggle comps. You need to determine where the test data comes from. In this comp, the research paper indicates that there is one big dataset and they split it between train and test. But in other comps, sometimes they use different datasets for test and train, and then you need to probe and watch LB more.", "votes": null }, { "id": "676331", "postDate": "11/19/2019 05:09:52", "content": "Biggest surprise for me was how a polluted CV worked just as well as a clean one. Each image is shared twice just a few hours apart. With 4 weeks to go I made 11-fold splits with 1 per year, monthly splits, you name it. No advantage over random 5-fold where the pairs were split train/valid.
", "rawMarkdown": "Biggest surprise for me was how a polluted CV worked just as well as a clean one. Each image is shared twice just a few hours apart. With 4 weeks to go I made 11-fold splits with 1 per year, monthly splits, you name it. No advantage over random 5-fold where the pairs were split train/valid.", "votes": null }, { "id": "676554", "postDate": "11/19/2019 09:49:02", "content": "My best private and public submission is the same one and it is the voting average of 5 best public LB submittions. I think it was the safe bet for this competition. It scored 0.65866 private, 0.66670 public.\nThe averaging method is:\n1. If mask is non-empty in at least 4 submissions, consider it non-empty.\n2. For non-empty mask, each pixel is 1 if it has at least two votes.
", "rawMarkdown": "My best private and public submission is the same one and it is the voting average of 5 best public LB submittions. I think it was the safe bet for this competition. It scored 0.65866 private, 0.66670 public.\nThe averaging method is:\n1. If mask is non-empty in at least 4 submissions, consider it non-empty.\n2. For non-empty mask, each pixel is 1 if it has at least two votes.", "votes": null } ], "comments": [ { "id": 676069, "author_name": "xiejialun", "author_url": "", "post_date": "11/19/2019 00:22:36", "content": "0.66154.. very close to my final submission. Big ensemble save me from the shake up Lol.\nBut shake up seems quite small in this competition.
", "votes": null, "replies": [] }, { "id": 676072, "author_name": "tugstugi", "author_url": "", "post_date": "11/19/2019 00:23:42", "content": "only 0.66104. fall down 115 places :) last day started to tune threshold...
", "votes": null, "replies": [ { "id": 676177, "author_name": "niuddd", "author_url": "", "post_date": "11/19/2019 02:21:04", "content": "Oh no, bad thing happens sometimes. And good luck next competition!
", "votes": null, "replies": [] } ] }, { "id": 676073, "author_name": "cweed28", "author_url": "", "post_date": "11/19/2019 00:25:51", "content": "0.65725. I used a different classifier
", "votes": null, "replies": [] }, { "id": 676077, "author_name": "mnpinto", "author_url": "", "post_date": "11/19/2019 00:26:52", "content": "private LB 0.66334, public LB 0.66883
", "votes": null, "replies": [] }, { "id": 676079, "author_name": "igormunizims", "author_url": "", "post_date": "11/19/2019 00:28:16", "content": "0.66824 (would be 9th place I think)\nthe interesting part is: this solution was an emsemble with models trained for each class that I thought was not working well
", "votes": null, "replies": [] }, { "id": 676083, "author_name": "ryches", "author_url": "", "post_date": "11/19/2019 00:31:20", "content": "My best submission was the one I picked. 20 small model (resnet18, resnet 34, efficientnetb0) ensemble with fixed max pixel threshold and fixed 0.3 mask threshold.
", "votes": null, "replies": [] }, { "id": 676103, "author_name": "mykttu", "author_url": "", "post_date": "11/19/2019 00:50:49", "content": "
tough one to digest. I'm sorry for your lost gold. Hope it will be a nice learning experience for you for future competitions
", "votes": null, "replies": [] }, { "id": 676180, "author_name": "niuddd", "author_url": "", "post_date": "11/19/2019 02:24:01", "content": "Lovazloss didn't work for me this time, and I hardly had success with inception, looking forward to your write up!
", "votes": null, "replies": [] } ] }, { "id": 676135, "author_name": "jionie", "author_url": "", "post_date": "11/19/2019 01:38:35", "content": "
Just tried one more submission after the competition ended.
", "votes": null, "replies": [] }, { "id": 676136, "author_name": "cdeotte", "author_url": "", "post_date": "11/19/2019 01:39:23", "content": "I learned my lesson from Steel comp and chose correctly this comp. I purposely chose less than my best public LB. I chose my best CV of 0.663 (with Public LB 0.670) which has Private LB 0.663. My best Public LB of 0.672 has Private LB 0.6615.
", "votes": null, "replies": [ { "id": 676144, "author_name": "phoenix9032", "author_url": "", "post_date": "11/19/2019 01:47:01", "content": "This competition , even though noisy data , the CV and LB correlated quite well ... So choosing based on CV was the right thing to do :) .. and no major shake up :D
", "votes": null, "replies": [] }, { "id": 676149, "author_name": "cdeotte", "author_url": "", "post_date": "11/19/2019 01:49:15", "content": "Yes. Congratulations Nijhar and team. You all did awesome!! What models did you use? I saw that you jumped many positions in the last few days.
", "votes": null, "replies": [] }, { "id": 676157, "author_name": "niuddd", "author_url": "", "post_date": "11/19/2019 01:59:31", "content": "This surprised me, bro. This is like the opposite of my situation...\nBut any reason about your determination?
", "votes": null, "replies": [] }, { "id": 676173, "author_name": "cdeotte", "author_url": "", "post_date": "11/19/2019 02:17:57", "content": "I had a feeling that my Public LB 0.672 model was overfitting. When my LB position was at 0.670, I did a few weird tricks and my score increased to LB 0.671. Then more tricks increased it to LB 0.672. But those same tricks didn't increase my CV, so I was suspicious that model was overfitting LB.
\n\nMy other model always had CV and LB change together.
", "votes": null, "replies": [] }, { "id": 676181, "author_name": "niuddd", "author_url": "", "post_date": "11/19/2019 02:25:53", "content": "Thanks. Trust your CV, lessons learned.
", "votes": null, "replies": [] }, { "id": 676208, "author_name": "phoenix9032", "author_url": "", "post_date": "11/19/2019 02:55:04", "content": "None of us actually thought it's possible to jump . But all credit goes to @kani23 . He was able to make Heng's model work and improve in last 2 days and we got a surprise improvement . When he gets up , he can explain first hand :)
", "votes": null, "replies": [] }, { "id": 676212, "author_name": "cdeotte", "author_url": "", "post_date": "11/19/2019 03:01:10", "content": "\n\n\nThanks. Trust your CV, lessons learned.
\n
Not always in Kaggle comps. You need to determine where the test data comes from. In this comp, the research paper indicates that there is one big dataset and they split it between train and test. But in other comps, sometimes they use different datasets for test and train, and then you need to probe and watch LB more.
", "votes": null, "replies": [] }, { "id": 676331, "author_name": "robga", "author_url": "", "post_date": "11/19/2019 05:09:52", "content": "Biggest surprise for me was how a polluted CV worked just as well as a clean one. Each image is shared twice just a few hours apart. With 4 weeks to go I made 11-fold splits with 1 per year, monthly splits, you name it. No advantage over random 5-fold where the pairs were split train/valid.
", "votes": null, "replies": [] } ] }, { "id": 676168, "author_name": "niuddd", "author_url": "", "post_date": "11/19/2019 02:12:25", "content": "OK this is me. At least we know up to 3 digits difference is untrustworthy.
\n\n
And I always used threshold=0.6 in the ensemble, that's about 38% positive, the only submission with lower threshold=0.55 gave best score. (I estimate it to be about 40% positive)
", "votes": null, "replies": [] } ] }, { "id": 676554, "author_name": "vzaguskin", "author_url": "", "post_date": "11/19/2019 09:49:02", "content": "My best private and public submission is the same one and it is the voting average of 5 best public LB submittions. I think it was the safe bet for this competition. It scored 0.65866 private, 0.66670 public.\nThe averaging method is:\n1. If mask is non-empty in at least 4 submissions, consider it non-empty.\n2. For non-empty mask, each pixel is 1 if it has at least two votes.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "676067": "Just curious...", "676069": "0.66154.. very close to my final submission. Big ensemble save me from the shake up Lol.\nBut shake up seems quite small in this competition.", "676072": "only 0.66104. fall down 115 places :) last day started to tune threshold...", "676073": "0.65725. I used a different classifier", "676077": "private LB 0.66334, public LB 0.66883", "676079": "0.66824 (would be 9th place I think)\nthe interesting part is: this solution was an emsemble with models trained for each class that I thought was not working well", "676083": "My best submission was the one I picked. 20 small model (resnet18, resnet 34, efficientnetb0) ensemble with fixed max pixel threshold and fixed 0.3 mask threshold.", "676103": "", "676114": "tough one to digest. I'm sorry for your lost gold. Hope it will be a nice learning experience for you for future competitions", "676135": "\n\nJust tried one more submission after the competition ended.", "676136": "I learned my lesson from Steel comp and chose correctly this comp. I purposely chose less than my best public LB. I chose my best CV of 0.663 (with Public LB 0.670) which has Private LB 0.663. My best Public LB of 0.672 has Private LB 0.6615.", "676144": "This competition , even though noisy data , the CV and LB correlated quite well ... So choosing based on CV was the right thing to do :) .. and no major shake up :D", "676149": "Yes. Congratulations Nijhar and team. You all did awesome!! What models did you use? I saw that you jumped many positions in the last few days.", "676157": "This surprised me, bro. This is like the opposite of my situation...\nBut any reason about your determination?", "676168": "OK this is me. At least we know up to 3 digits difference is untrustworthy.\n\n", "676171": "And I always used threshold=0.6 in the ensemble, that's about 38% positive, the only submission with lower threshold=0.55 gave best score. (I estimate it to be about 40% positive)", "676173": "I had a feeling that my Public LB 0.672 model was overfitting. When my LB position was at 0.670, I did a few weird tricks and my score increased to LB 0.671. Then more tricks increased it to LB 0.672. But those same tricks didn't increase my CV, so I was suspicious that model was overfitting LB.\n\nMy other model always had CV and LB change together.", "676177": "Oh no, bad thing happens sometimes. And good luck next competition!", "676180": "Lovazloss didn't work for me this time, and I hardly had success with inception, looking forward to your write up!", "676181": "Thanks. Trust your CV, lessons learned.", "676208": "None of us actually thought it's possible to jump . But all credit goes to @kani23 . He was able to make Heng's model work and improve in last 2 days and we got a surprise improvement . When he gets up , he can explain first hand :)", "676212": "> Thanks. Trust your CV, lessons learned.\n\nNot always in Kaggle comps. You need to determine where the test data comes from. In this comp, the research paper indicates that there is one big dataset and they split it between train and test. But in other comps, sometimes they use different datasets for test and train, and then you need to probe and watch LB more.", "676331": "Biggest surprise for me was how a polluted CV worked just as well as a clean one. Each image is shared twice just a few hours apart. With 4 weeks to go I made 11-fold splits with 1 per year, monthly splits, you name it. No advantage over random 5-fold where the pairs were split train/valid.", "676554": "My best private and public submission is the same one and it is the voting average of 5 best public LB submittions. I think it was the safe bet for this competition. It scored 0.65866 private, 0.66670 public.\nThe averaging method is:\n1. If mask is non-empty in at least 4 submissions, consider it non-empty.\n2. For non-empty mask, each pixel is 1 if it has at least two votes." }, "source": "meta" }