Has anyone tried combining the metadata features with the image data? Was it hard to do? Any improvements of your model?
", "messages": [ { "id": "214684", "postDate": "08/17/2017 20:44:14", "content": "Has anyone tried combining the metadata features with the image data? Was it hard to do? Any improvements of your model?
", "rawMarkdown": "Has anyone tried combining the metadata features with the image data? Was it hard to do? Any improvements of your model?", "votes": null }, { "id": "214790", "postDate": "08/18/2017 09:57:14", "content": "As far as I know - main reason for LB 99.7 instead of 99.9 is Error in the masks and shadows\nDo not think that either of those could be fixed via metadata
", "rawMarkdown": "As far as I know - main reason for LB 99.7 instead of 99.9 is Error in the masks and shadows\nDo not think that either of those could be fixed via metadata", "votes": null }, { "id": "214807", "postDate": "08/18/2017 11:10:31", "content": "how have you came up to this conclusion? I mean LB is close to CV. If you fix train masks and retrain CCN and compare train scores - they have still the same order.
", "rawMarkdown": "how have you came up to this conclusion? I mean LB is close to CV. If you fix train masks and retrain CCN and compare train scores - they have still the same order.", "votes": null }, { "id": "214895", "postDate": "08/18/2017 17:38:29", "content": "I have no delusions that 0.999 is reachable, but I thought that extra information could be useful for going from, say, 0.9981 to 0.9982. Seeing that the top of the leaderboard is starting to saturate with really high scores, I thought this might be one of the possible extra pieces of data worth exploring. And since local validation seems to be fairly reliable in this competition, I thought people might be able to tell if the metafeatures are helpful.
", "rawMarkdown": "I have no delusions that 0.999 is reachable, but I thought that extra information could be useful for going from, say, 0.9981 to 0.9982. Seeing that the top of the leaderboard is starting to saturate with really high scores, I thought this might be one of the possible extra pieces of data worth exploring. And since local validation seems to be fairly reliable in this competition, I thought people might be able to tell if the metafeatures are helpful.", "votes": null }, { "id": "214919", "postDate": "08/18/2017 19:25:01", "content": "I have 0.996 LB score and when I look at my validation test outputs, it becomes obvious that there is a lot of room for improvement. I think 0.999 is reachable. There is not that much noise in the train masks.
", "rawMarkdown": "I have 0.996 LB score and when I look at my validation test outputs, it becomes obvious that there is a lot of room for improvement. I think 0.999 is reachable. There is not that much noise in the train masks.", "votes": null }, { "id": "214947", "postDate": "08/18/2017 21:12:30", "content": "There is a thread with corrupted masks from train set - you cans estimate its proportion based on this
", "rawMarkdown": "There is a thread with corrupted masks from train set - you cans estimate its proportion based on this", "votes": null }, { "id": "214961", "postDate": "08/18/2017 23:43:47", "content": "What are the metadata features?
", "rawMarkdown": "What are the metadata features?", "votes": null }, { "id": "214984", "postDate": "08/19/2017 03:28:07", "content": "Yeah, probably no more than 100 images with less than 1% mistake in average. I might be wrong though.
", "rawMarkdown": "Yeah, probably no more than 100 images with less than 1% mistake in average. I might be wrong though.", "votes": null }, { "id": "215008", "postDate": "08/19/2017 07:19:51", "content": "There is about a 100 images with mistakes bog enough to notice it from the first glance - there are much more smaller mistakes and we are already struggling for 0.25% from the total \nSo it could be the case
", "rawMarkdown": "There is about a 100 images with mistakes bog enough to notice it from the first glance - there are much more smaller mistakes and we are already struggling for 0.25% from the total \nSo it could be the case", "votes": null }, { "id": "215046", "postDate": "08/19/2017 11:46:43", "content": "Hi @Bojan Tunguz
\n\nI have been thinking about using Concat or Merge of keras to incorporate the \"differences between consecutive image\", is this what you meant by extra information?
", "rawMarkdown": "Hi @Bojan Tunguz\n\nI have been thinking about using Concat or Merge of keras to incorporate the \"differences between consecutive image\", is this what you meant by extra information?", "votes": null } ], "comments": [ { "id": 214790, "author_name": "venheads", "author_url": "", "post_date": "08/18/2017 09:57:14", "content": "As far as I know - main reason for LB 99.7 instead of 99.9 is Error in the masks and shadows\nDo not think that either of those could be fixed via metadata
", "votes": null, "replies": [ { "id": 214807, "author_name": "heyt0ny", "author_url": "", "post_date": "08/18/2017 11:10:31", "content": "how have you came up to this conclusion? I mean LB is close to CV. If you fix train masks and retrain CCN and compare train scores - they have still the same order.
", "votes": null, "replies": [] }, { "id": 214919, "author_name": "harungunaydin", "author_url": "", "post_date": "08/18/2017 19:25:01", "content": "I have 0.996 LB score and when I look at my validation test outputs, it becomes obvious that there is a lot of room for improvement. I think 0.999 is reachable. There is not that much noise in the train masks.
", "votes": null, "replies": [] }, { "id": 214947, "author_name": "venheads", "author_url": "", "post_date": "08/18/2017 21:12:30", "content": "There is a thread with corrupted masks from train set - you cans estimate its proportion based on this
", "votes": null, "replies": [] }, { "id": 214984, "author_name": "harungunaydin", "author_url": "", "post_date": "08/19/2017 03:28:07", "content": "Yeah, probably no more than 100 images with less than 1% mistake in average. I might be wrong though.
", "votes": null, "replies": [] }, { "id": 215008, "author_name": "venheads", "author_url": "", "post_date": "08/19/2017 07:19:51", "content": "There is about a 100 images with mistakes bog enough to notice it from the first glance - there are much more smaller mistakes and we are already struggling for 0.25% from the total \nSo it could be the case
", "votes": null, "replies": [] } ] }, { "id": 214895, "author_name": "tunguz", "author_url": "", "post_date": "08/18/2017 17:38:29", "content": "I have no delusions that 0.999 is reachable, but I thought that extra information could be useful for going from, say, 0.9981 to 0.9982. Seeing that the top of the leaderboard is starting to saturate with really high scores, I thought this might be one of the possible extra pieces of data worth exploring. And since local validation seems to be fairly reliable in this competition, I thought people might be able to tell if the metafeatures are helpful.
", "votes": null, "replies": [ { "id": 215046, "author_name": "janpreets", "author_url": "", "post_date": "08/19/2017 11:46:43", "content": "Hi @Bojan Tunguz
\n\nI have been thinking about using Concat or Merge of keras to incorporate the \"differences between consecutive image\", is this what you meant by extra information?
", "votes": null, "replies": [] } ] }, { "id": 214961, "author_name": "xiaokangwang", "author_url": "", "post_date": "08/18/2017 23:43:47", "content": "What are the metadata features?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "214684": "Has anyone tried combining the metadata features with the image data? Was it hard to do? Any improvements of your model?", "214790": "As far as I know - main reason for LB 99.7 instead of 99.9 is Error in the masks and shadows\nDo not think that either of those could be fixed via metadata", "214807": "how have you came up to this conclusion? I mean LB is close to CV. If you fix train masks and retrain CCN and compare train scores - they have still the same order.", "214895": "I have no delusions that 0.999 is reachable, but I thought that extra information could be useful for going from, say, 0.9981 to 0.9982. Seeing that the top of the leaderboard is starting to saturate with really high scores, I thought this might be one of the possible extra pieces of data worth exploring. And since local validation seems to be fairly reliable in this competition, I thought people might be able to tell if the metafeatures are helpful.", "214919": "I have 0.996 LB score and when I look at my validation test outputs, it becomes obvious that there is a lot of room for improvement. I think 0.999 is reachable. There is not that much noise in the train masks.", "214947": "There is a thread with corrupted masks from train set - you cans estimate its proportion based on this", "214961": "What are the metadata features?", "214984": "Yeah, probably no more than 100 images with less than 1% mistake in average. I might be wrong though.", "215008": "There is about a 100 images with mistakes bog enough to notice it from the first glance - there are much more smaller mistakes and we are already struggling for 0.25% from the total \nSo it could be the case", "215046": "Hi @Bojan Tunguz\n\nI have been thinking about using Concat or Merge of keras to incorporate the \"differences between consecutive image\", is this what you meant by extra information?" }, "source": "meta" }