Thanks to everyone that took part in the competition this year. We were very impressed by the results overall!
\n\nI would like to encourage participants to share their approaches in this thread e.g. what worked/what didn't work.
\n\nThanks again and well done!
", "messages": [ { "id": 550627, "postDate": "2019-06-11T20:58:56.543Z", "content": "Thanks to everyone that took part in the competition this year. We were very impressed by the results overall!
\n\nI would like to encourage participants to share their approaches in this thread e.g. what worked/what didn't work.
\n\nThanks again and well done!
", "rawMarkdown": "Thanks to everyone that took part in the competition this year. We were very impressed by the results overall!\n\nI would like to encourage participants to share their approaches in this thread e.g. what worked/what didn't work. \n\nThanks again and well done!", "votes": 3 }, { "id": 565396, "postDate": "2019-06-30T22:20:42.097Z", "content": "Oisin posted the solution slides from the winners (except the top team) on the workshop page. (solutions)
\n\nThank you Oisin and the winners! Didn't really have time to compete, but it's very interesting to learn the approaches.
", "rawMarkdown": "Oisin posted the solution slides from the winners (except the top team) on the workshop page. ([solutions](https://drive.google.com/file/d/1T8XABp6ayFTRpwbEYPFYICfIRXrD1fJB/view))\n\nThank you Oisin and the winners! Didn't really have time to compete, but it's very interesting to learn the approaches." }, { "id": 555377, "postDate": "2019-06-18T20:29:58.483Z", "content": "resnet34 and resnet50 pretrained on imagenet didn't really work. maybe not enough epochs, i could only do 3-4 on kaggle kernels due to time limit. only got score just below 0.5.
", "rawMarkdown": "resnet34 and resnet50 pretrained on imagenet didn't really work. maybe not enough epochs, i could only do 3-4 on kaggle kernels due to time limit. only got score just below 0.5.", "replies": [ { "id": 555547, "postDate": "2019-06-19T03:59:36.550Z", "content": "Those are pretty shallow nets, you'll probably find they aren't sufficient for getting state-of-the-art results on this dataset. Also, as you observed, 4 epochs is really not enough. Even if you have a well tuned super-convergence (https://arxiv.org/abs/1708.07120) configuration, I doubt you'll get very far with only 4 epochs. 30-40 epochs at a minimum, and for the absolute best performance, I would plan on more like 100 epochs.
", "rawMarkdown": "Those are pretty shallow nets, you'll probably find they aren't sufficient for getting state-of-the-art results on this dataset. Also, as you observed, 4 epochs is really not enough. Even if you have a well tuned super-convergence (https://arxiv.org/abs/1708.07120) configuration, I doubt you'll get very far with only 4 epochs. 30-40 epochs at a minimum, and for the absolute best performance, I would plan on more like 100 epochs.", "votes": 1 } ] }, { "id": 551557, "postDate": "2019-06-12T21:12:30.407Z", "content": "Teams are typically removed after the end of a competition if there is evidence of multiple accounts or other team irregularities. After checking with your team you can reach out to https://www.kaggle.com/compliance with questions.
", "rawMarkdown": "Teams are typically removed after the end of a competition if there is evidence of multiple accounts or other team irregularities. After checking with your team you can reach out to https://www.kaggle.com/compliance with questions. " }, { "id": 550775, "postDate": "2019-06-12T02:44:02.050Z", "content": "Could you please help me find out that why our team were removed? QAQ
", "rawMarkdown": "Could you please help me find out that why our team were removed? QAQ" }, { "id": 550963, "postDate": "2019-06-12T07:22:45.487Z", "rawMarkdown": "", "isDeleted": true }, { "id": 550912, "postDate": "2019-06-12T06:16:49.730Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 565396, "author_name": "Naoto Usuyama", "author_url": "", "post_date": "2019-06-30T22:20:42.097000", "content": "Oisin posted the solution slides from the winners (except the top team) on the workshop page. (solutions)
\n\nThank you Oisin and the winners! Didn't really have time to compete, but it's very interesting to learn the approaches.
", "votes": 0, "replies": [] }, { "id": 555377, "author_name": "miwojc", "author_url": "", "post_date": "2019-06-18T20:29:58.483000", "content": "resnet34 and resnet50 pretrained on imagenet didn't really work. maybe not enough epochs, i could only do 3-4 on kaggle kernels due to time limit. only got score just below 0.5.
", "votes": 0, "replies": [ { "id": 555547, "author_name": "Number Six", "author_url": "", "post_date": "2019-06-19T03:59:36.550000", "content": "Those are pretty shallow nets, you'll probably find they aren't sufficient for getting state-of-the-art results on this dataset. Also, as you observed, 4 epochs is really not enough. Even if you have a well tuned super-convergence (https://arxiv.org/abs/1708.07120) configuration, I doubt you'll get very far with only 4 epochs. 30-40 epochs at a minimum, and for the absolute best performance, I would plan on more like 100 epochs.
", "votes": 1, "replies": [] } ] }, { "id": 551557, "author_name": "Maggie", "author_url": "", "post_date": "2019-06-12T21:12:30.407000", "content": "Teams are typically removed after the end of a competition if there is evidence of multiple accounts or other team irregularities. After checking with your team you can reach out to https://www.kaggle.com/compliance with questions.
", "votes": 0, "replies": [] }, { "id": 550775, "author_name": "LiuXun", "author_url": "", "post_date": "2019-06-12T02:44:02.050000", "content": "Could you please help me find out that why our team were removed? QAQ
", "votes": 0, "replies": [] }, { "id": 550963, "author_name": "", "author_url": "", "post_date": "2019-06-12T07:22:45.487000", "content": "", "votes": 0, "replies": [] }, { "id": 550912, "author_name": "", "author_url": "", "post_date": "2019-06-12T06:16:49.730000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "550627": "Thanks to everyone that took part in the competition this year. We were very impressed by the results overall!\n\nI would like to encourage participants to share their approaches in this thread e.g. what worked/what didn't work. \n\nThanks again and well done!", "565396": "Oisin posted the solution slides from the winners (except the top team) on the workshop page. ([solutions](https://drive.google.com/file/d/1T8XABp6ayFTRpwbEYPFYICfIRXrD1fJB/view))\n\nThank you Oisin and the winners! Didn't really have time to compete, but it's very interesting to learn the approaches.", "555377": "resnet34 and resnet50 pretrained on imagenet didn't really work. maybe not enough epochs, i could only do 3-4 on kaggle kernels due to time limit. only got score just below 0.5.", "551557": "Teams are typically removed after the end of a competition if there is evidence of multiple accounts or other team irregularities. After checking with your team you can reach out to https://www.kaggle.com/compliance with questions. ", "550775": "Could you please help me find out that why our team were removed? QAQ", "550963": "", "550912": "" } }