After training multiple model for more than 2 weeks and bad submission results, Today I realize that post processing is the key. Thanks @cdeotte, I got this hint after reading topics you posted in this competition.
", "messages": [ { "id": "671758", "postDate": "11/13/2019 05:55:40", "content": "After training multiple model for more than 2 weeks and bad submission results, Today I realize that post processing is the key. Thanks @cdeotte, I got this hint after reading topics you posted in this competition.
", "rawMarkdown": "After training multiple model for more than 2 weeks and bad submission results, Today I realize that post processing is the key. Thanks @cdeotte, I got this hint after reading topics you posted in this competition.", "votes": null }, { "id": "671764", "postDate": "11/13/2019 06:05:52", "content": "Wonderful. If you tweak your model (architecture and loss) you can have it output masks without the need for post processing, but I have gotten better results by not restricting my networks too much and instead using post processing too.
", "rawMarkdown": "Wonderful. If you tweak your model (architecture and loss) you can have it output masks without the need for post processing, but I have gotten better results by not restricting my networks too much and instead using post processing too.", "votes": null }, { "id": "671780", "postDate": "11/13/2019 06:32:27", "content": "Use post processing to remove false positives. Eg by size or by max pixel probability
", "rawMarkdown": "Use post processing to remove false positives. Eg by size or by max pixel probability", "votes": null }, { "id": "671886", "postDate": "11/13/2019 09:19:12", "content": "I won't call it the key..\nIt is quite risky to keep tuning the parameter of post-processing for just increasing the public score.\nIn fact, I have kept trying to get rid of post-processing since last week.. But it is pretty hard though..\nBut still, good luck with you
", "rawMarkdown": "I won't call it the key..\nIt is quite risky to keep tuning the parameter of post-processing for just increasing the public score.\nIn fact, I have kept trying to get rid of post-processing since last week.. But it is pretty hard though..\nBut still, good luck with you", "votes": null }, { "id": "671924", "postDate": "11/13/2019 10:28:09", "content": "Thanks @hengck23, I tried mask by size.. now i will try max pixel probability and best of luck for you!!!!
", "rawMarkdown": "Thanks @hengck23, I tried mask by size.. now i will try max pixel probability and best of luck for you!!!!", "votes": null }, { "id": "672106", "postDate": "11/13/2019 14:15:35", "content": "@xiejialun I did not tuned any parameter. I just used 3 lines of post processing code from Chris notebook thereby got boost of 0.003X on public LB and my rank improved from top 15% to top 8%. Then I realized that post processing may be key.
", "rawMarkdown": "xiejialun I did not tuned any parameter. I just used 3 lines of post processing code from Chris notebook thereby got boost of 0.003X on public LB and my rank improved from top 15% to top 8%. Then I realized that post processing may be key.", "votes": null }, { "id": "672616", "postDate": "11/14/2019 02:44:52", "content": "I think post-processing is very useful until LB score < 0.670 😂
", "rawMarkdown": "I think post-processing is very useful until LB score < 0.670 😂", "votes": null }, { "id": "672621", "postDate": "11/14/2019 02:51:05", "content": "I see, since I drop a lot of score in steel competition with threshold tuning. Lol\nGood luck with you! I'm sure yours is fine.
", "rawMarkdown": "I see, since I drop a lot of score in steel competition with threshold tuning. Lol\nGood luck with you! I'm sure yours is fine.", "votes": null }, { "id": "672622", "postDate": "11/14/2019 02:53:06", "content": "If you don't mind, do you mean the pixel threshold part, or the mask size threshold part? Since I do feel I can't get more benefit from mask size threshold. But pixel threshold still impact my score a lot.
", "rawMarkdown": "If you don't mind, do you mean the pixel threshold part, or the mask size threshold part? Since I do feel I can't get more benefit from mask size threshold. But pixel threshold still impact my score a lot.", "votes": null }, { "id": "672654", "postDate": "11/14/2019 03:41:50", "content": "I did mean mask_size 😃 (mask.sum() < 1000)
", "rawMarkdown": "I did mean mask_size 😃 (mask.sum() < 1000)", "votes": null }, { "id": "672663", "postDate": "11/14/2019 03:53:08", "content": "wow.. 1000 is quite small. Mine is 10000 😂.\nYour model did predict quite accurate, good luck with you!
", "rawMarkdown": "wow.. 1000 is quite small. Mine is 10000 😂.\nYour model did predict quite accurate, good luck with you!", "votes": null }, { "id": "672677", "postDate": "11/14/2019 04:09:52", "content": "(mask.sum() < 1000) is just an example. 😂 \nI didn’t use that kind of post-processing. \nGood luck with you, too!
", "rawMarkdown": "(mask.sum() < 1000) is just an example. 😂 \nI didn’t use that kind of post-processing. \nGood luck with you, too!", "votes": null }, { "id": "673023", "postDate": "11/14/2019 12:12:10", "content": "For classify empty mask, I have tried two logic, I think in fact they do the same thing:\n1. if mask.sum()<20000, then mask=0\n2. max_pool(logit_layer)<0.5, then mask=0
", "rawMarkdown": "For classify empty mask, I have tried two logic, I think in fact they do the same thing:\n1. if mask.sum()<20000, then mask=0\n2. max_pool(logit_layer)<0.5, then mask=0", "votes": null }, { "id": "673048", "postDate": "11/14/2019 12:55:05", "content": "Yeah post-processing is the key but first you need a good model I guess
", "rawMarkdown": "Yeah post-processing is the key but first you need a good model I guess", "votes": null } ], "comments": [ { "id": 671764, "author_name": "cdeotte", "author_url": "", "post_date": "11/13/2019 06:05:52", "content": "Wonderful. If you tweak your model (architecture and loss) you can have it output masks without the need for post processing, but I have gotten better results by not restricting my networks too much and instead using post processing too.
", "votes": null, "replies": [] }, { "id": 671780, "author_name": "hengck23", "author_url": "", "post_date": "11/13/2019 06:32:27", "content": "Use post processing to remove false positives. Eg by size or by max pixel probability
", "votes": null, "replies": [ { "id": 671924, "author_name": "raghaw", "author_url": "", "post_date": "11/13/2019 10:28:09", "content": "Thanks @hengck23, I tried mask by size.. now i will try max pixel probability and best of luck for you!!!!
", "votes": null, "replies": [] } ] }, { "id": 671886, "author_name": "xiejialun", "author_url": "", "post_date": "11/13/2019 09:19:12", "content": "I won't call it the key..\nIt is quite risky to keep tuning the parameter of post-processing for just increasing the public score.\nIn fact, I have kept trying to get rid of post-processing since last week.. But it is pretty hard though..\nBut still, good luck with you
", "votes": null, "replies": [ { "id": 672106, "author_name": "raghaw", "author_url": "", "post_date": "11/13/2019 14:15:35", "content": "@xiejialun I did not tuned any parameter. I just used 3 lines of post processing code from Chris notebook thereby got boost of 0.003X on public LB and my rank improved from top 15% to top 8%. Then I realized that post processing may be key.
", "votes": null, "replies": [] }, { "id": 672621, "author_name": "xiejialun", "author_url": "", "post_date": "11/14/2019 02:51:05", "content": "I see, since I drop a lot of score in steel competition with threshold tuning. Lol\nGood luck with you! I'm sure yours is fine.
", "votes": null, "replies": [] } ] }, { "id": 672616, "author_name": "limerobot", "author_url": "", "post_date": "11/14/2019 02:44:52", "content": "I think post-processing is very useful until LB score < 0.670 😂
", "votes": null, "replies": [ { "id": 672622, "author_name": "xiejialun", "author_url": "", "post_date": "11/14/2019 02:53:06", "content": "If you don't mind, do you mean the pixel threshold part, or the mask size threshold part? Since I do feel I can't get more benefit from mask size threshold. But pixel threshold still impact my score a lot.
", "votes": null, "replies": [] }, { "id": 672654, "author_name": "limerobot", "author_url": "", "post_date": "11/14/2019 03:41:50", "content": "I did mean mask_size 😃 (mask.sum() < 1000)
", "votes": null, "replies": [] }, { "id": 672663, "author_name": "xiejialun", "author_url": "", "post_date": "11/14/2019 03:53:08", "content": "wow.. 1000 is quite small. Mine is 10000 😂.\nYour model did predict quite accurate, good luck with you!
", "votes": null, "replies": [] }, { "id": 672677, "author_name": "limerobot", "author_url": "", "post_date": "11/14/2019 04:09:52", "content": "(mask.sum() < 1000) is just an example. 😂 \nI didn’t use that kind of post-processing. \nGood luck with you, too!
", "votes": null, "replies": [] } ] }, { "id": 673023, "author_name": "niuddd", "author_url": "", "post_date": "11/14/2019 12:12:10", "content": "For classify empty mask, I have tried two logic, I think in fact they do the same thing:\n1. if mask.sum()<20000, then mask=0\n2. max_pool(logit_layer)<0.5, then mask=0
", "votes": null, "replies": [] }, { "id": 673048, "author_name": "silverstone1903", "author_url": "", "post_date": "11/14/2019 12:55:05", "content": "Yeah post-processing is the key but first you need a good model I guess
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "671758": "After training multiple model for more than 2 weeks and bad submission results, Today I realize that post processing is the key. Thanks @cdeotte, I got this hint after reading topics you posted in this competition.", "671764": "Wonderful. If you tweak your model (architecture and loss) you can have it output masks without the need for post processing, but I have gotten better results by not restricting my networks too much and instead using post processing too.", "671780": "Use post processing to remove false positives. Eg by size or by max pixel probability", "671886": "I won't call it the key..\nIt is quite risky to keep tuning the parameter of post-processing for just increasing the public score.\nIn fact, I have kept trying to get rid of post-processing since last week.. But it is pretty hard though..\nBut still, good luck with you", "671924": "Thanks @hengck23, I tried mask by size.. now i will try max pixel probability and best of luck for you!!!!", "672106": "xiejialun I did not tuned any parameter. I just used 3 lines of post processing code from Chris notebook thereby got boost of 0.003X on public LB and my rank improved from top 15% to top 8%. Then I realized that post processing may be key.", "672616": "I think post-processing is very useful until LB score < 0.670 😂", "672621": "I see, since I drop a lot of score in steel competition with threshold tuning. Lol\nGood luck with you! I'm sure yours is fine.", "672622": "If you don't mind, do you mean the pixel threshold part, or the mask size threshold part? Since I do feel I can't get more benefit from mask size threshold. But pixel threshold still impact my score a lot.", "672654": "I did mean mask_size 😃 (mask.sum() < 1000)", "672663": "wow.. 1000 is quite small. Mine is 10000 😂.\nYour model did predict quite accurate, good luck with you!", "672677": "(mask.sum() < 1000) is just an example. 😂 \nI didn’t use that kind of post-processing. \nGood luck with you, too!", "673023": "For classify empty mask, I have tried two logic, I think in fact they do the same thing:\n1. if mask.sum()<20000, then mask=0\n2. max_pool(logit_layer)<0.5, then mask=0", "673048": "Yeah post-processing is the key but first you need a good model I guess" }, "source": "meta" }