I was thinking of implementing Siamese Network after trying many ways of classification. But wanted to know how much well it is performing for this task. If you've received better results with it as compared to classification, could you please let us know and what all changes may be beneficial ?
", "messages": [ { "id": 445815, "postDate": "2018-12-27T04:59:21.523Z", "content": "I was thinking of implementing Siamese Network after trying many ways of classification. But wanted to know how much well it is performing for this task. If you've received better results with it as compared to classification, could you please let us know and what all changes may be beneficial ?
", "rawMarkdown": "I was thinking of implementing Siamese Network after trying many ways of classification. But wanted to know how much well it is performing for this task. If you've received better results with it as compared to classification, could you please let us know and what all changes may be beneficial ? ", "votes": 3 }, { "id": 447676, "postDate": "2018-12-30T10:05:53.380Z", "content": "I think it can easily get you 0.9+ with some fine-tuning. In fact, I expect to see 0.98+ submission all over the LB in the later stage of this competition.
", "rawMarkdown": "I think it can easily get you 0.9+ with some fine-tuning. In fact, I expect to see 0.98+ submission all over the LB in the later stage of this competition.", "votes": 1, "replies": [ { "id": 448015, "postDate": "2018-12-31T03:58:47.537Z", "content": "You might be true, although on discussions, seems like it's not giving comparable results as compared to mere classification. Also how do you handle new_whale?
I have not tried any model in this competition. I believe that Siamese can give you over 0.85LB.
I took part in the playground competition before. Also, I know exactly some guys are doing Siamese network and they archive >0.85LB
My single Siamese got 0.897 with only 90 epochs compared to the playground solution's 0.822 after 400.
\n\nAs for the reason why a siamese is good, because this is a one-shot detection problem, not classification. A really good classification model may work, but it's just much easier to do it the way it's intended to.
", "rawMarkdown": "My single Siamese got 0.897 with only 90 epochs compared to the playground solution's 0.822 after 400.\n\nAs for the reason why a siamese is good, because this is a one-shot detection problem, not classification. A really good classification model *may* work, but it's just much easier to do it the way it's intended to.", "votes": 6 }, { "id": 452214, "postDate": "2019-01-08T11:15:05.563Z", "content": "Achieving 0.897 by only single model is wonderful. I think you can improve more 0.01~0.02 by ensembling multiple models.
", "rawMarkdown": "Achieving 0.897 by only single model is wonderful. I think you can improve more 0.01~0.02 by ensembling multiple models." }, { "id": 452220, "postDate": "2019-01-08T11:22:17.070Z", "content": "You can get 0.9 with classification only. Although, I also think that metric learning is more appropriate for this task
", "rawMarkdown": "You can get 0.9 with classification only. Although, I also think that metric learning is more appropriate for this task", "votes": 6 }, { "id": 452735, "postDate": "2019-01-09T05:12:52.697Z", "content": "0.9 with classification only ? That's great. I tried classification with many models, doesn't seem to get above 0.78. Which model do you seem to be using, is it custom ?
", "rawMarkdown": "0.9 with classification only ? That's great. I tried classification with many models, doesn't seem to get above 0.78. Which model do you seem to be using, is it custom ?" }, { "id": 457374, "postDate": "2019-01-17T09:50:01.387Z", "content": "It is heavily based on https://www.kaggle.com/c/humpback-whale-identification/discussion/74647 . That is why my nickname for this competition is \"thanks radek\"
", "rawMarkdown": "It is heavily based on https://www.kaggle.com/c/humpback-whale-identification/discussion/74647 . That is why my nickname for this competition is \"thanks radek\"", "votes": 3 }, { "id": 467385, "postDate": "2019-02-07T02:06:49.130Z", "content": "Hi, old-ufo, Which notebook of radek's is your solution based on? Thank you.
", "rawMarkdown": "Hi, old-ufo, Which notebook of radek's is your solution based on? Thank you." }, { "id": 467525, "postDate": "2019-02-07T09:18:44.613Z", "content": "https://github.com/radekosmulski/whale/blob/master/only_known_train.ipynb
", "rawMarkdown": "https://github.com/radekosmulski/whale/blob/master/only_known_train.ipynb", "votes": 1 }, { "id": 467724, "postDate": "2019-02-07T16:14:01.137Z", "content": "Nice, Thanks for your sharing.
", "rawMarkdown": "Nice, Thanks for your sharing." }, { "id": 469533, "postDate": "2019-02-11T11:33:26.177Z", "content": "Thanks for reading this question, I got one thing that I don't understand, how to solve the unbalanced distribution problem if you treat this problem as a Classification task? Thanks again for your time
", "rawMarkdown": "Thanks for reading this question, I got one thing that I don't understand, how to solve the unbalanced distribution problem if you treat this problem as a Classification task? Thanks again for your time" } ] } ], "comments": [ { "id": 447676, "author_name": "Khoi Nguyen", "author_url": "", "post_date": "2018-12-30T10:05:53.380000", "content": "I think it can easily get you 0.9+ with some fine-tuning. In fact, I expect to see 0.98+ submission all over the LB in the later stage of this competition.
", "votes": 1, "replies": [ { "id": 448015, "author_name": "Prajjwal", "author_url": "", "post_date": "2018-12-31T03:58:47.537000", "content": "You might be true, although on discussions, seems like it's not giving comparable results as compared to mere classification. Also how do you handle new_whale?
I have not tried any model in this competition. I believe that Siamese can give you over 0.85LB.
I took part in the playground competition before. Also, I know exactly some guys are doing Siamese network and they archive >0.85LB
My single Siamese got 0.897 with only 90 epochs compared to the playground solution's 0.822 after 400.
\n\nAs for the reason why a siamese is good, because this is a one-shot detection problem, not classification. A really good classification model may work, but it's just much easier to do it the way it's intended to.
", "votes": 6, "replies": [] }, { "id": 452214, "author_name": "toshi_k", "author_url": "", "post_date": "2019-01-08T11:15:05.563000", "content": "Achieving 0.897 by only single model is wonderful. I think you can improve more 0.01~0.02 by ensembling multiple models.
", "votes": 0, "replies": [] }, { "id": 452220, "author_name": "old-ufo", "author_url": "", "post_date": "2019-01-08T11:22:17.070000", "content": "You can get 0.9 with classification only. Although, I also think that metric learning is more appropriate for this task
", "votes": 6, "replies": [] }, { "id": 452735, "author_name": "Prajjwal", "author_url": "", "post_date": "2019-01-09T05:12:52.697000", "content": "0.9 with classification only ? That's great. I tried classification with many models, doesn't seem to get above 0.78. Which model do you seem to be using, is it custom ?
", "votes": 0, "replies": [] }, { "id": 457374, "author_name": "old-ufo", "author_url": "", "post_date": "2019-01-17T09:50:01.387000", "content": "It is heavily based on https://www.kaggle.com/c/humpback-whale-identification/discussion/74647 . That is why my nickname for this competition is \"thanks radek\"
", "votes": 3, "replies": [] }, { "id": 467385, "author_name": "Gary", "author_url": "", "post_date": "2019-02-07T02:06:49.130000", "content": "Hi, old-ufo, Which notebook of radek's is your solution based on? Thank you.
", "votes": 0, "replies": [] }, { "id": 467525, "author_name": "old-ufo", "author_url": "", "post_date": "2019-02-07T09:18:44.613000", "content": "https://github.com/radekosmulski/whale/blob/master/only_known_train.ipynb
", "votes": 1, "replies": [] }, { "id": 467724, "author_name": "Gary", "author_url": "", "post_date": "2019-02-07T16:14:01.137000", "content": "Nice, Thanks for your sharing.
", "votes": 0, "replies": [] }, { "id": 469533, "author_name": "TheRedArrow", "author_url": "", "post_date": "2019-02-11T11:33:26.177000", "content": "Thanks for reading this question, I got one thing that I don't understand, how to solve the unbalanced distribution problem if you treat this problem as a Classification task? Thanks again for your time
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "445815": "I was thinking of implementing Siamese Network after trying many ways of classification. But wanted to know how much well it is performing for this task. If you've received better results with it as compared to classification, could you please let us know and what all changes may be beneficial ? ", "447676": "I think it can easily get you 0.9+ with some fine-tuning. In fact, I expect to see 0.98+ submission all over the LB in the later stage of this competition.", "445831": "I have not tried any model in this competition. I believe that `Siamese` can give you over `0.85LB`." } }