Just a bunch of different models\nresnet34 36 tiles, resnet34 49 tiles, resnet50 36 tiles, b3 36 tiles, b5 25 tiles, b0 36 tiles and different tiles sizes (256 and 128) + D4 TTA gave .933. Indeed, diverse models are the key :) \nUnfortunately, we picked the best CV-holdout submit. Worked pretty well, but still :(
", "messages": [ { "id": 940442, "postDate": "2020-07-23T00:29:27.440Z", "content": "missed 0.941 single model 4 fold( but it is rather noise than miss, and one got more luck with selection of the final sub than others. Though, all subs using this particular approach got 0.935-0.941 while were performing about the same on public LB as my main approach I have chosen.
", "rawMarkdown": "**missed 0.941 single model 4 fold**( but it is rather noise than miss, and one got more luck with selection of the final sub than others. Though, all subs using this particular approach got 0.935-0.941 while were performing about the same on public LB as my main approach I have chosen.", "votes": 6, "replies": [ { "id": 940449, "postDate": "2020-07-23T00:34:00.333Z", "content": "Great thanks lafoss!\nYour approach gave us a lot of improvements.\nI am waiting your final solution!
", "rawMarkdown": "Great thanks lafoss!\nYour approach gave us a lot of improvements.\nI am waiting your final solution!", "votes": 3 }, { "id": 940456, "postDate": "2020-07-23T00:41:56.190Z", "content": "You are welcome, I'm working on it right now
", "rawMarkdown": "You are welcome, I'm working on it right now", "votes": 2 }, { "id": 940461, "postDate": "2020-07-23T00:49:28.640Z", "content": "Well @iafoss I can't wait to see your pipeline as well. You've had a consistent lead on this competition, put out there a lot of great content, while helping a lot of people who were asking you questions sometimes even unrelated to your work. The consistency of your results seems to imply that not everything was random in this competition. I think at this point, consistency of scores has more value that a third decimal. I'm sorry you didn't get the money ^^
", "rawMarkdown": "Well @iafoss I can't wait to see your pipeline as well. You've had a consistent lead on this competition, put out there a lot of great content, while helping a lot of people who were asking you questions sometimes even unrelated to your work. The consistency of your results seems to imply that not everything was random in this competition. I think at this point, consistency of scores has more value that a third decimal. I'm sorry you didn't get the money ^^", "votes": 2 }, { "id": 940588, "postDate": "2020-07-23T03:27:43.683Z", "content": "Thank you for your pipeline :) I think it helped lots of people to get started!
", "rawMarkdown": "Thank you for your pipeline :) I think it helped lots of people to get started!", "votes": 1 }, { "id": 942397, "postDate": "2020-07-23T18:11:18.407Z", "content": "You are welcome.
", "rawMarkdown": "You are welcome." } ] }, { "id": 940425, "postDate": "2020-07-23T00:10:31.777Z", "content": "Just a bunch of different models\nresnet34 36 tiles, resnet34 49 tiles, resnet50 36 tiles, b3 36 tiles, b5 25 tiles, b0 36 tiles and different tiles sizes (256 and 128) + D4 TTA gave .933. Indeed, diverse models are the key :) \nUnfortunately, we picked the best CV-holdout submit. Worked pretty well, but still :(
", "rawMarkdown": "Just a bunch of different models\nresnet34 36 tiles, resnet34 49 tiles, resnet50 36 tiles, b3 36 tiles, b5 25 tiles, b0 36 tiles and different tiles sizes (256 and 128) + D4 TTA gave .933. Indeed, diverse models are the key :) \nUnfortunately, we picked the best CV-holdout submit. Worked pretty well, but still :( ", "votes": 5 }, { "id": 940447, "postDate": "2020-07-23T00:31:41.940Z", "content": "Congrats, your hard work (I can’t experiments such many models) and thanks sharing fir your approach! \nYes.. I also missed .924 scoring model..\nBut I learned a lot of things from this competition.
", "rawMarkdown": "Congrats, your hard work (I can’t experiments such many models) and thanks sharing fir your approach! \nYes.. I also missed .924 scoring model..\nBut I learned a lot of things from this competition.", "votes": 1 }, { "id": 940428, "postDate": "2020-07-23T00:17:22.123Z", "content": "We also missed .927 scoring model. Anyways congratulations and thanks for sharing your approach.
", "rawMarkdown": "We also missed .927 scoring model. Anyways congratulations and thanks for sharing your approach.", "votes": 1 }, { "id": 940545, "postDate": "2020-07-23T02:36:34.617Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 940442, "author_name": "Iafoss", "author_url": "", "post_date": "2020-07-23T00:29:27.440000", "content": "missed 0.941 single model 4 fold( but it is rather noise than miss, and one got more luck with selection of the final sub than others. Though, all subs using this particular approach got 0.935-0.941 while were performing about the same on public LB as my main approach I have chosen.
", "votes": 6, "replies": [ { "id": 940449, "author_name": "Nishi yasu", "author_url": "", "post_date": "2020-07-23T00:34:00.333000", "content": "Great thanks lafoss!\nYour approach gave us a lot of improvements.\nI am waiting your final solution!
", "votes": 3, "replies": [] }, { "id": 940456, "author_name": "Iafoss", "author_url": "", "post_date": "2020-07-23T00:41:56.190000", "content": "You are welcome, I'm working on it right now
", "votes": 2, "replies": [] }, { "id": 940461, "author_name": "Benjamin Dubreu", "author_url": "", "post_date": "2020-07-23T00:49:28.640000", "content": "Well @iafoss I can't wait to see your pipeline as well. You've had a consistent lead on this competition, put out there a lot of great content, while helping a lot of people who were asking you questions sometimes even unrelated to your work. The consistency of your results seems to imply that not everything was random in this competition. I think at this point, consistency of scores has more value that a third decimal. I'm sorry you didn't get the money ^^
", "votes": 2, "replies": [] }, { "id": 940588, "author_name": "kvigly", "author_url": "", "post_date": "2020-07-23T03:27:43.683000", "content": "Thank you for your pipeline :) I think it helped lots of people to get started!
", "votes": 1, "replies": [] }, { "id": 942397, "author_name": "Iafoss", "author_url": "", "post_date": "2020-07-23T18:11:18.407000", "content": "You are welcome.
", "votes": 0, "replies": [] } ] }, { "id": 940447, "author_name": "Nishi yasu", "author_url": "", "post_date": "2020-07-23T00:31:41.940000", "content": "Congrats, your hard work (I can’t experiments such many models) and thanks sharing fir your approach! \nYes.. I also missed .924 scoring model..\nBut I learned a lot of things from this competition.
", "votes": 1, "replies": [] }, { "id": 940428, "author_name": "Karan", "author_url": "", "post_date": "2020-07-23T00:17:22.123000", "content": "We also missed .927 scoring model. Anyways congratulations and thanks for sharing your approach.
", "votes": 1, "replies": [] }, { "id": 940545, "author_name": "", "author_url": "", "post_date": "2020-07-23T02:36:34.617000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "940442": "**missed 0.941 single model 4 fold**( but it is rather noise than miss, and one got more luck with selection of the final sub than others. Though, all subs using this particular approach got 0.935-0.941 while were performing about the same on public LB as my main approach I have chosen.", "940425": "Just a bunch of different models\nresnet34 36 tiles, resnet34 49 tiles, resnet50 36 tiles, b3 36 tiles, b5 25 tiles, b0 36 tiles and different tiles sizes (256 and 128) + D4 TTA gave .933. Indeed, diverse models are the key :) \nUnfortunately, we picked the best CV-holdout submit. Worked pretty well, but still :( ", "940447": "Congrats, your hard work (I can’t experiments such many models) and thanks sharing fir your approach! \nYes.. I also missed .924 scoring model..\nBut I learned a lot of things from this competition.", "940428": "We also missed .927 scoring model. Anyways congratulations and thanks for sharing your approach.", "940545": "" } }