{ "id": 274925, "title": "分享一下code。private LB 0.98681，public LB 0.98159 late submission", "url": "/competitions/classify-leaves/discussion/274925", "author_name": "", "post_date": "2021-09-28T06:26:08.573634100Z", "votes": 12, "comment_count": 4, "views": 0, "content": "

比赛结束这么久了，不知道还有没有人看了😂

\n

分享一下自己的code https://github.com/SHENZHENYI/Classify-Leaves-Kaggle-MuLi-d2l-course

\n

code是在自己的baseline上加了@seefun的一些tricks。

\n

最终结果是private LB 0.98681，public LB 0.98159，不过参与太晚是late submission所以LB上没名字。
\n这次比赛经历收获很大（学习到了很多干货label smoothing, mixup, timm的model zoo, AdamW, mixed-precision training, TTA, 5fold投票）。个人还发现PIL和cv2的resize会影响结果，cv2的好像普遍差点（实验了三次），最后把cv2.Resize中的interpolation改成interpolation=cv2.INTER_AREA貌似把问题解决了。

\n

谢谢大家在Discussion和Code区中的分享，也谢谢李沐开设d2l的课程和这场比赛，希望课程办的越来越好。

", "messages": [ { "id": "1526478", "postDate": "09/28/2021 06:26:08", "content": "

比赛结束这么久了，不知道还有没有人看了😂

\n

分享一下自己的code https://github.com/SHENZHENYI/Classify-Leaves-Kaggle-MuLi-d2l-course

\n

code是在自己的baseline上加了@seefun的一些tricks。

\n

最终结果是private LB 0.98681，public LB 0.98159，不过参与太晚是late submission所以LB上没名字。
\n这次比赛经历收获很大（学习到了很多干货label smoothing, mixup, timm的model zoo, AdamW, mixed-precision training, TTA, 5fold投票）。个人还发现PIL和cv2的resize会影响结果，cv2的好像普遍差点（实验了三次），最后把cv2.Resize中的interpolation改成interpolation=cv2.INTER_AREA貌似把问题解决了。

\n

谢谢大家在Discussion和Code区中的分享，也谢谢李沐开设d2l的课程和这场比赛，希望课程办的越来越好。

", "rawMarkdown": "比赛结束这么久了，不知道还有没有人看了😂\n\n分享一下自己的code https://github.com/SHENZHENYI/Classify-Leaves-Kaggle-MuLi-d2l-course\n\ncode是在自己的baseline上加了@seefun的一些tricks。\n\n最终结果是private LB 0.98681，public LB 0.98159，不过参与太晚是late submission所以LB上没名字。\n这次比赛经历收获很大（学习到了很多干货label smoothing, mixup, timm的model zoo, AdamW, mixed-precision training, TTA, 5fold投票）。个人还发现PIL和cv2的resize会影响结果，cv2的好像普遍差点（实验了三次），最后把cv2.Resize中的interpolation改成interpolation=cv2.INTER_AREA貌似把问题解决了。\n\n谢谢大家在Discussion和Code区中的分享，也谢谢李沐开设d2l的课程和这场比赛，希望课程办的越来越好。", "votes": null }, { "id": "1574127", "postDate": "11/07/2021 09:13:30", "content": "

感谢分享！
\n我看了一下你的源代码，发现了一些小问题，请问下面的train_csv = csv.iloc[train_folds[2]].reset_index()里的2是不是应该改为fold呢？

\n

for fold in range(FOLD):
\nprint(f'Start Fold{fold}...')
\n train_csv = csv.iloc[train_folds[2]].reset_index()
\n val_csv = csv.iloc[val_folds[2]].reset_index()

", "rawMarkdown": "感谢分享！\n我看了一下你的源代码，发现了一些小问题，请问下面的`train_csv = csv.iloc[train_folds[2]].reset_index()`里的2是不是应该改为fold呢？\n\n`for fold in range(FOLD):`\n` print(f'Start Fold{fold}...')`\n ` train_csv = csv.iloc[train_folds[2]].reset_index()`\n ` val_csv = csv.iloc[val_folds[2]].reset_index()`", "votes": null }, { "id": "1578403", "postDate": "11/11/2021 01:43:31", "content": "

对，应该是fold。我当时为了跑快点，实验时候把FOLD=1，所以在train_folds里乱写了个数字😷。感谢指正！！！

", "rawMarkdown": "对，应该是fold。我当时为了跑快点，实验时候把FOLD=1，所以在train_folds里乱写了个数字😷。感谢指正！！！", "votes": null }, { "id": "1749171", "postDate": "04/08/2022 09:40:32", "content": "

May I know the 'Find the LR' part code is coming from what trick?

\n

Why do we need 'num = len(train_dl) - 1' minus 1 for the dataloader batch?

\n

Thanks.

", "rawMarkdown": "May I know the 'Find the LR' part code is coming from what trick? \n\nWhy do we need 'num = len(train_dl) - 1' minus 1 for the dataloader batch? \n\nThanks.", "votes": null }, { "id": "2195892", "postDate": "03/25/2023 00:21:42", "content": "

您好，请问您代码里的'clean_train_v4.csv'是什么文件啊，给的数据集并没有这个文件啊

", "rawMarkdown": "您好，请问您代码里的'clean_train_v4.csv'是什么文件啊，给的数据集并没有这个文件啊", "votes": null } ], "comments": [ { "id": 1574127, "author_name": "wubba1ubadubdub", "author_url": "", "post_date": "11/07/2021 09:13:30", "content": "

感谢分享！
\n我看了一下你的源代码，发现了一些小问题，请问下面的train_csv = csv.iloc[train_folds[2]].reset_index()里的2是不是应该改为fold呢？

\n

for fold in range(FOLD):
\nprint(f'Start Fold{fold}...')
\n train_csv = csv.iloc[train_folds[2]].reset_index()
\n val_csv = csv.iloc[val_folds[2]].reset_index()

", "votes": null, "replies": [ { "id": 1578403, "author_name": "zhenyishen", "author_url": "", "post_date": "11/11/2021 01:43:31", "content": "

对，应该是fold。我当时为了跑快点，实验时候把FOLD=1，所以在train_folds里乱写了个数字😷。感谢指正！！！

", "votes": null, "replies": [] } ] }, { "id": 1749171, "author_name": "brandontey", "author_url": "", "post_date": "04/08/2022 09:40:32", "content": "

May I know the 'Find the LR' part code is coming from what trick?

\n

Why do we need 'num = len(train_dl) - 1' minus 1 for the dataloader batch?

\n

Thanks.

", "votes": null, "replies": [] }, { "id": 2195892, "author_name": "meiyuxin", "author_url": "", "post_date": "03/25/2023 00:21:42", "content": "

您好，请问您代码里的'clean_train_v4.csv'是什么文件啊，给的数据集并没有这个文件啊

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1526478": "比赛结束这么久了，不知道还有没有人看了😂\n\n分享一下自己的code https://github.com/SHENZHENYI/Classify-Leaves-Kaggle-MuLi-d2l-course\n\ncode是在自己的baseline上加了@seefun的一些tricks。\n\n最终结果是private LB 0.98681，public LB 0.98159，不过参与太晚是late submission所以LB上没名字。\n这次比赛经历收获很大（学习到了很多干货label smoothing, mixup, timm的model zoo, AdamW, mixed-precision training, TTA, 5fold投票）。个人还发现PIL和cv2的resize会影响结果，cv2的好像普遍差点（实验了三次），最后把cv2.Resize中的interpolation改成interpolation=cv2.INTER_AREA貌似把问题解决了。\n\n谢谢大家在Discussion和Code区中的分享，也谢谢李沐开设d2l的课程和这场比赛，希望课程办的越来越好。", "1574127": "感谢分享！\n我看了一下你的源代码，发现了一些小问题，请问下面的`train_csv = csv.iloc[train_folds[2]].reset_index()`里的2是不是应该改为fold呢？\n\n`for fold in range(FOLD):`\n` print(f'Start Fold{fold}...')`\n ` train_csv = csv.iloc[train_folds[2]].reset_index()`\n ` val_csv = csv.iloc[val_folds[2]].reset_index()`", "1578403": "对，应该是fold。我当时为了跑快点，实验时候把FOLD=1，所以在train_folds里乱写了个数字😷。感谢指正！！！", "1749171": "May I know the 'Find the LR' part code is coming from what trick? \n\nWhy do we need 'num = len(train_dl) - 1' minus 1 for the dataloader batch? \n\nThanks.", "2195892": "您好，请问您代码里的'clean_train_v4.csv'是什么文件啊，给的数据集并没有这个文件啊" }, "source": "meta" }