![\"sub\"](\"https://z3.ax1x.com/2021/06/18/RSlu9g.png\")
I choose two sub with public LB 0.97772 and 0.98613. (my best LB score is 0.98795)
\nI don't choose the best submission because I believe too high score in this noisy labeled dataset may suffer from over-fitting.
I choose two sub with public LB 0.97772 and 0.98613. (my best LB score is 0.98795)
\nI don't choose the best submission because I believe too high score in this noisy labeled dataset may suffer from over-fitting.
请问你觉得大概什么样的结果算是比较好的呢，顺便分享一下你的思路呗
", "rawMarkdown": "请问你觉得大概什么样的结果算是比较好的呢，顺便分享一下你的思路呗", "votes": null }, { "id": "1356416", "postDate": "06/19/2021 02:36:36", "content": "我的个人经验是:
\nupdate:
\nIn last two competitions, I submitted the wrong submission twice , both of the time I followed somebody else's method and I thought that I wasn't good enough to research and study and I shouldn't submit my method and mine method was the best submission both of the times, I got so poor private score than expected ,after these two competitions I released that we should respect our own work too, like I am doing kaggle continuously form last two months and this is the the lesson I have got that everyone should believe in their own working methodology more than others.
\nAnd @seefun well done man, that you have this skill, excellent mate.
", "rawMarkdown": "*In last two competitions, I submitted the wrong submission twice , both of the time I followed somebody else's method and I thought that I wasn't good enough to research and study and I shouldn't submit my method and mine method was the best submission both of the times, I got so poor private score than expected ,after these two competitions I released that we should respect our own work too, like I am doing kaggle continuously form last two months and this is the the lesson I have got that everyone should believe in their own working methodology more than others.*\n\nAnd @seefun well done man, that you have this skill, excellent mate.", "votes": null }, { "id": "1365877", "postDate": "06/26/2021 08:00:36", "content": "翻车了 应该是测试集里太多重复样本和噪声标注了
", "rawMarkdown": "翻车了 应该是测试集里太多重复样本和噪声标注了", "votes": null }, { "id": "1367800", "postDate": "06/28/2021 05:19:12", "content": "哈哈哈相信沐神没错了，ResNeSt真的效果极佳😄
", "rawMarkdown": "哈哈哈相信沐神没错了，ResNeSt真的效果极佳😄", "votes": null } ], "comments": [ { "id": 1355583, "author_name": "crazywinds", "author_url": "", "post_date": "06/18/2021 11:45:19", "content": "请问你觉得大概什么样的结果算是比较好的呢，顺便分享一下你的思路呗
", "votes": null, "replies": [ { "id": 1356416, "author_name": "seefun", "author_url": "", "post_date": "06/19/2021 02:36:36", "content": "我的个人经验是:
\nupdate:
\n哈哈哈相信沐神没错了，ResNeSt真的效果极佳😄
", "votes": null, "replies": [] } ] }, { "id": 1362883, "author_name": "sayantankirtaniya", "author_url": "", "post_date": "06/23/2021 18:01:02", "content": "In last two competitions, I submitted the wrong submission twice , both of the time I followed somebody else's method and I thought that I wasn't good enough to research and study and I shouldn't submit my method and mine method was the best submission both of the times, I got so poor private score than expected ,after these two competitions I released that we should respect our own work too, like I am doing kaggle continuously form last two months and this is the the lesson I have got that everyone should believe in their own working methodology more than others.
\nAnd @seefun well done man, that you have this skill, excellent mate.
", "votes": null, "replies": [] }, { "id": 1365877, "author_name": "seefun", "author_url": "", "post_date": "06/26/2021 08:00:36", "content": "翻车了 应该是测试集里太多重复样本和噪声标注了
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1354953": "I choose two sub with public LB 0.97772 and 0.98613. (my best LB score is 0.98795)\nI don't choose the best submission because I believe too high score in this noisy labeled dataset may suffer from over-fitting.\n\n", "1355583": "请问你觉得大概什么样的结果算是比较好的呢，顺便分享一下你的思路呗", "1356416": "我的个人经验是:\n1. 选取**合适的验证方法**。比如AUC，mAP(mean average precision)，micro/macro F-beta score 等，根据任务和数据分布选择最适合的验证方法。一般来说，ACC(Precision)是较为不可靠的验证指标，尤其是正负样本不均衡的情况下(比如正负样本9:1,全预测1也有90%分)。\n2. **相信本地交叉验证**结果而不是线上LB。因为线上测试数据量更小，而交叉验证数据量却是整个训练数据集。\n3. **相信自己**。一些降低模型预测方差的方法，如TTA（test-time augmentation），Blending等集成方法，能降低shake风险（防止public LB分很高而private LB分很低）。一些自己的经验相信有效的方法，不要因为不稳定的测试集而轻易放弃。\n4. 对于CV任务，**优秀的特征提取**更为重要。使用metric learning等方法可以获取更好的特征提取，这样的模型通常会更鲁棒更值得相信。（目前我只提交了baseline，还尚未尝试这点）。\n\n---------\n\nupdate:\n5. **相信沐神**。[Bag of Tricks](https://arxiv.org/pdf/1812.01187.pdf) | [ResNeSt](https://arxiv.org/abs/2004.08955)", "1362883": "*In last two competitions, I submitted the wrong submission twice , both of the time I followed somebody else's method and I thought that I wasn't good enough to research and study and I shouldn't submit my method and mine method was the best submission both of the times, I got so poor private score than expected ,after these two competitions I released that we should respect our own work too, like I am doing kaggle continuously form last two months and this is the the lesson I have got that everyone should believe in their own working methodology more than others.*\n\nAnd @seefun well done man, that you have this skill, excellent mate.", "1365877": "翻车了 应该是测试集里太多重复样本和噪声标注了", "1367800": "哈哈哈相信沐神没错了，ResNeSt真的效果极佳😄" }, "source": "meta" }