It’s over! It’s over!
", "messages": [ { "id": "315253", "postDate": "04/17/2018 00:02:37", "content": "It’s over! It’s over!
", "rawMarkdown": "It’s over! It’s over!", "votes": null }, { "id": "315254", "postDate": "04/17/2018 00:03:18", "content": "Not till the LB updates.
", "rawMarkdown": "Not till the LB updates.", "votes": null }, { "id": "315269", "postDate": "04/17/2018 00:30:53", "content": "Allen, What is your private LB score? You made us run like madness with your incredible stage1 score...
", "rawMarkdown": "Allen, What is your private LB score? You made us run like madness with your incredible stage1 score...", "votes": null }, { "id": "315282", "postDate": "04/17/2018 00:41:45", "content": "My best model scored a 0.582. My top model from phase one was in the 40s, but that was intentionally (and aggressively) overfit to find the upper-bound. However, all the usual caveats apply. Unfortunately, I can’t say whether this good or bad for a few days! I’ve also started merging code from the competition back into Keras-RCNN. I plan on getting everything merged in the next month or two.
", "rawMarkdown": "My best model scored a 0.582. My top model from phase one was in the 40s, but that was intentionally (and aggressively) overfit to find the upper-bound. However, all the usual caveats apply. Unfortunately, I can’t say whether this good or bad for a few days! I’ve also started merging code from the competition back into [Keras-RCNN][1]. I plan on getting everything merged in the next month or two. \n\n\n [1]: https://github.com/broadinstitute/keras-rcnn", "votes": null }, { "id": "315581", "postDate": "04/17/2018 09:59:28", "content": "@Allen Goodman thanks for all your contributions.
", "rawMarkdown": "Allen Goodman thanks for all your contributions.", "votes": null }, { "id": "315595", "postDate": "04/17/2018 10:29:41", "content": "It's very impressive the winning solution scored so close to your result on the stage 1 LB while doing so well on the new unseen data of the stage2.
", "rawMarkdown": "It's very impressive the winning solution scored so close to your result on the stage 1 LB while doing so well on the new unseen data of the stage2.", "votes": null }, { "id": "315723", "postDate": "04/17/2018 14:20:02", "content": "So after we have seen that the winning team used a u-net, do you thing a Mask-RCNN could have performed as good or is u-net the way to go for the next similar competition?
", "rawMarkdown": "So after we have seen that the winning team used a u-net, do you thing a Mask-RCNN could have performed as good or is u-net the way to go for the next similar competition?", "votes": null }, { "id": "315730", "postDate": "04/17/2018 14:24:56", "content": "I’ve said this elsewhere, but I think it was really a trade-off between minimizing localization or segmentation error. I don’t know about future competitions, but I’d expect multi-class is the future and I don’t know whether fully convolutional networks will continue to do well. I guess we’ll find out. 👍
", "rawMarkdown": "I’ve said this elsewhere, but I think it was really a trade-off between minimizing localization or segmentation error. I don’t know about future competitions, but I’d expect multi-class is the future and I don’t know whether fully convolutional networks will continue to do well. I guess we’ll find out. 👍", "votes": null }, { "id": "315732", "postDate": "04/17/2018 14:26:02", "content": "I agree! I feel like I managed to find the upper bound but for six teams (!!!) with six different solutions to achieve that (or close to that) on both phases was super impressive.
", "rawMarkdown": "I agree! I feel like I managed to find the upper bound but for six teams (!!!) with six different solutions to achieve that (or close to that) on both phases was super impressive.", "votes": null }, { "id": "315745", "postDate": "04/17/2018 14:37:54", "content": "I'm really impressed that stage 2 scores were so high!
", "rawMarkdown": "I'm really impressed that stage 2 scores were so high!", "votes": null }, { "id": "315971", "postDate": "04/17/2018 22:53:47", "content": "Hi @Anne Carpenter, is what we are seeing right now on the private LB the stage2 scores? They looked the same as the stage1 for us so I am not sure.
", "rawMarkdown": "Hi @Anne Carpenter, is what we are seeing right now on the private LB the stage2 scores? They looked the same as the stage1 for us so I am not sure.", "votes": null }, { "id": "316621", "postDate": "04/19/2018 14:16:45", "content": "As far as I know, yes. I now see an orange banner on the private LB saying the results are not final until verified by Kaggle.
", "rawMarkdown": "As far as I know, yes. I now see an orange banner on the private LB saying the results are not final until verified by Kaggle.", "votes": null } ], "comments": [ { "id": 315254, "author_name": "dredata", "author_url": "", "post_date": "04/17/2018 00:03:18", "content": "Not till the LB updates.
", "votes": null, "replies": [] }, { "id": 315269, "author_name": "cpmpml", "author_url": "", "post_date": "04/17/2018 00:30:53", "content": "Allen, What is your private LB score? You made us run like madness with your incredible stage1 score...
", "votes": null, "replies": [] }, { "id": 315282, "author_name": "agoodman", "author_url": "", "post_date": "04/17/2018 00:41:45", "content": "My best model scored a 0.582. My top model from phase one was in the 40s, but that was intentionally (and aggressively) overfit to find the upper-bound. However, all the usual caveats apply. Unfortunately, I can’t say whether this good or bad for a few days! I’ve also started merging code from the competition back into Keras-RCNN. I plan on getting everything merged in the next month or two.
", "votes": null, "replies": [ { "id": 315595, "author_name": "dmytropoplavskiy", "author_url": "", "post_date": "04/17/2018 10:29:41", "content": "It's very impressive the winning solution scored so close to your result on the stage 1 LB while doing so well on the new unseen data of the stage2.
", "votes": null, "replies": [] }, { "id": 315732, "author_name": "agoodman", "author_url": "", "post_date": "04/17/2018 14:26:02", "content": "I agree! I feel like I managed to find the upper bound but for six teams (!!!) with six different solutions to achieve that (or close to that) on both phases was super impressive.
", "votes": null, "replies": [] }, { "id": 315745, "author_name": "drannecarpenter", "author_url": "", "post_date": "04/17/2018 14:37:54", "content": "I'm really impressed that stage 2 scores were so high!
", "votes": null, "replies": [] }, { "id": 315971, "author_name": "sheriytm", "author_url": "", "post_date": "04/17/2018 22:53:47", "content": "Hi @Anne Carpenter, is what we are seeing right now on the private LB the stage2 scores? They looked the same as the stage1 for us so I am not sure.
", "votes": null, "replies": [] }, { "id": 316621, "author_name": "drannecarpenter", "author_url": "", "post_date": "04/19/2018 14:16:45", "content": "As far as I know, yes. I now see an orange banner on the private LB saying the results are not final until verified by Kaggle.
", "votes": null, "replies": [] } ] }, { "id": 315581, "author_name": "sheriytm", "author_url": "", "post_date": "04/17/2018 09:59:28", "content": "@Allen Goodman thanks for all your contributions.
", "votes": null, "replies": [] }, { "id": 315723, "author_name": "christophneuner", "author_url": "", "post_date": "04/17/2018 14:20:02", "content": "So after we have seen that the winning team used a u-net, do you thing a Mask-RCNN could have performed as good or is u-net the way to go for the next similar competition?
", "votes": null, "replies": [ { "id": 315730, "author_name": "agoodman", "author_url": "", "post_date": "04/17/2018 14:24:56", "content": "I’ve said this elsewhere, but I think it was really a trade-off between minimizing localization or segmentation error. I don’t know about future competitions, but I’d expect multi-class is the future and I don’t know whether fully convolutional networks will continue to do well. I guess we’ll find out. 👍
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "315253": "It’s over! It’s over!", "315254": "Not till the LB updates.", "315269": "Allen, What is your private LB score? You made us run like madness with your incredible stage1 score...", "315282": "My best model scored a 0.582. My top model from phase one was in the 40s, but that was intentionally (and aggressively) overfit to find the upper-bound. However, all the usual caveats apply. Unfortunately, I can’t say whether this good or bad for a few days! I’ve also started merging code from the competition back into [Keras-RCNN][1]. I plan on getting everything merged in the next month or two. \n\n\n [1]: https://github.com/broadinstitute/keras-rcnn", "315581": "Allen Goodman thanks for all your contributions.", "315595": "It's very impressive the winning solution scored so close to your result on the stage 1 LB while doing so well on the new unseen data of the stage2.", "315723": "So after we have seen that the winning team used a u-net, do you thing a Mask-RCNN could have performed as good or is u-net the way to go for the next similar competition?", "315730": "I’ve said this elsewhere, but I think it was really a trade-off between minimizing localization or segmentation error. I don’t know about future competitions, but I’d expect multi-class is the future and I don’t know whether fully convolutional networks will continue to do well. I guess we’ll find out. 👍", "315732": "I agree! I feel like I managed to find the upper bound but for six teams (!!!) with six different solutions to achieve that (or close to that) on both phases was super impressive.", "315745": "I'm really impressed that stage 2 scores were so high!", "315971": "Hi @Anne Carpenter, is what we are seeing right now on the private LB the stage2 scores? They looked the same as the stage1 for us so I am not sure.", "316621": "As far as I know, yes. I now see an orange banner on the private LB saying the results are not final until verified by Kaggle." }, "source": "meta" }