Congrats to everyone and host of competition!\nI am surprised that I am get 3thd place on private leaderboard. I was just lucky 🙂 . I stopped improving my solution 1 month ago, because has problems with my GPU and didn't want to spend more credits on AWS, because I have not seen any improvement on LB and CV.
\n\nI am experimented with different networks and my custom tile cropping, hard augmentations. But the best results I get on simple sollution based on @haqishen kernel.\nI am trained 2 effnetb0 and used round logits before average them, this approach give me 0.880 on public and 0.934 on private.
\n\nGithub with my experiments: https://github.com/Dipet/kaggle_panda
", "messages": [ { "id": 941540, "postDate": "2020-07-23T09:35:25.860Z", "content": "Congrats to everyone and host of competition!\nI am surprised that I am get 3thd place on private leaderboard. I was just lucky 🙂 . I stopped improving my solution 1 month ago, because has problems with my GPU and didn't want to spend more credits on AWS, because I have not seen any improvement on LB and CV.
\n\nI am experimented with different networks and my custom tile cropping, hard augmentations. But the best results I get on simple sollution based on @haqishen kernel.\nI am trained 2 effnetb0 and used round logits before average them, this approach give me 0.880 on public and 0.934 on private.
\n\nGithub with my experiments: https://github.com/Dipet/kaggle_panda
", "rawMarkdown": "Congrats to everyone and host of competition!\nI am surprised that I am get 3thd place on private leaderboard. I was just lucky 🙂 . I stopped improving my solution 1 month ago, because has problems with my GPU and didn't want to spend more credits on AWS, because I have not seen any improvement on LB and CV.\n\nI am experimented with different networks and my custom tile cropping, hard augmentations. But the best results I get on simple sollution based on @haqishen [kernel](https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87).\nI am trained 2 effnetb0 and used round logits before average them, this approach give me 0.880 on public and 0.934 on private.\n\nGithub with my experiments: https://github.com/Dipet/kaggle_panda", "votes": 22 }, { "id": 942820, "postDate": "2020-07-24T02:56:54.670Z", "content": "Great going ! Congratulations.
", "rawMarkdown": "Great going ! Congratulations. " }, { "id": 941743, "postDate": "2020-07-23T11:40:24.560Z", "content": "congratulations!👍
", "rawMarkdown": "congratulations!👍 " } ], "comments": [ { "id": 942820, "author_name": "jwelliav", "author_url": "", "post_date": "2020-07-24T02:56:54.670000", "content": "Great going ! Congratulations.
", "votes": 0, "replies": [] }, { "id": 941743, "author_name": "Deemo Chen", "author_url": "", "post_date": "2020-07-23T11:40:24.560000", "content": "congratulations!👍
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "941540": "Congrats to everyone and host of competition!\nI am surprised that I am get 3thd place on private leaderboard. I was just lucky 🙂 . I stopped improving my solution 1 month ago, because has problems with my GPU and didn't want to spend more credits on AWS, because I have not seen any improvement on LB and CV.\n\nI am experimented with different networks and my custom tile cropping, hard augmentations. But the best results I get on simple sollution based on @haqishen [kernel](https://www.kaggle.com/haqishen/train-efficientnet-b0-w-36-tiles-256-lb0-87).\nI am trained 2 effnetb0 and used round logits before average them, this approach give me 0.880 on public and 0.934 on private.\n\nGithub with my experiments: https://github.com/Dipet/kaggle_panda", "942820": "Great going ! Congratulations. ", "941743": "congratulations!👍 " } }