This challenge is a fine-grained image recognition task, so siamese is a popular solution.\nIn the previous competition Humpback Whale Identification, we used Siamese and it works well, I know the whale task is difference from iMet, graphs in Whale dataset are actually very similar, but iMet have much classes that each class is not similar.\nSiamese need to use more computing power and takes long time to train.\nso I want to ask if you have you tried to use siamese in this competition and if it works well?
", "messages": [ { "id": 518552, "postDate": "2019-04-17T12:50:16.963Z", "content": "Each picture has more than one labels. So I don't think triplet loss or any other metric learning method works well in this competition without some tricks like label clustering.
", "rawMarkdown": "Each picture has more than one labels. So I don't think triplet loss or any other metric learning method works well in this competition without some tricks like label clustering.", "votes": 3, "replies": [ { "id": 518599, "postDate": "2019-04-17T13:57:34.187Z", "content": "Hmmmm, may I ask why do you think triplet loss won't work well if we can one hot encoded the labels?
", "rawMarkdown": "Hmmmm, may I ask why do you think triplet loss won't work well if we can one hot encoded the labels?", "votes": 2 }, { "id": 518601, "postDate": "2019-04-17T14:00:24.057Z", "content": "It 's just my guess. I think it should be verified in experiment.
", "rawMarkdown": "It 's just my guess. I think it should be verified in experiment.", "votes": 3 }, { "id": 518624, "postDate": "2019-04-17T14:31:47.993Z", "content": "I think it’s a good idea to encode the labels, let’s try it
", "rawMarkdown": "I think it’s a good idea to encode the labels, let’s try it", "votes": 1 }, { "id": 520475, "postDate": "2019-04-21T04:24:23.810Z", "content": "Though I haven't try, some loss functions like triplet loss + label clustering would be good. I think we cannot use triplet loss + one-hot encoding directly, some sub-classes should be correlated like (American) and (American or European). One-hot encoding would consider them as independent.
", "rawMarkdown": "Though I haven't try, some loss functions like triplet loss + label clustering would be good. I think we cannot use triplet loss + one-hot encoding directly, some sub-classes should be correlated like (American) and (American or European). One-hot encoding would consider them as independent. ", "votes": 1 } ] }, { "id": 513205, "postDate": "2019-04-11T03:28:57.897Z", "content": "May not , Siamese network is a similarity measure method, which can be used to identify and classify categories when there are many categories but the number of samples per category is small. And the similarity of iMet dataset is not high as humpback whale identification.
", "rawMarkdown": "May not , Siamese network is a similarity measure method, which can be used to identify and classify categories when there are many categories but the number of samples per category is small. And the similarity of iMet dataset is not high as humpback whale identification.", "votes": 3, "replies": [ { "id": 513232, "postDate": "2019-04-11T03:43:31.380Z", "content": "Thank you
", "rawMarkdown": "Thank you" } ] }, { "id": 508498, "postDate": "2019-04-06T08:49:00.783Z", "content": "This challenge is a fine-grained image recognition task, so siamese is a popular solution.\nIn the previous competition Humpback Whale Identification, we used Siamese and it works well, I know the whale task is difference from iMet, graphs in Whale dataset are actually very similar, but iMet have much classes that each class is not similar.\nSiamese need to use more computing power and takes long time to train.\nso I want to ask if you have you tried to use siamese in this competition and if it works well?
", "rawMarkdown": "This challenge is a fine-grained image recognition task, so siamese is a popular solution.\nIn the previous competition [Humpback Whale Identification](https://www.kaggle.com/c/humpback-whale-identification), we used Siamese and it works well, I know the whale task is difference from iMet, graphs in Whale dataset are actually very similar, but iMet have much classes that each class is not similar.\nSiamese need to use more computing power and takes long time to train.\nso I want to ask if you have you tried to use siamese in this competition and if it works well?", "votes": 3 }, { "id": 516334, "postDate": "2019-04-14T02:45:00.657Z", "content": "Siamese network is quite powerful, I don’t see why it would not be successful here. But as others mentioned, it takes a very long time to train. Also, using the mpiotte Siamese net that was popular in whales will be tricky to use here without modifications, there are too many images to use exact lap solutions as was done there. You could try approximate lap. Not sure yet but classification seems like the way to start here.
", "rawMarkdown": "Siamese network is quite powerful, I don’t see why it would not be successful here. But as others mentioned, it takes a very long time to train. Also, using the mpiotte Siamese net that was popular in whales will be tricky to use here without modifications, there are too many images to use exact lap solutions as was done there. You could try approximate lap. Not sure yet but classification seems like the way to start here. ", "votes": 1 }, { "id": 509345, "postDate": "2019-04-07T17:30:26.967Z", "content": "I am actually struggling with some of the bigger pretrained models (getting them to complete with reasonable param settings in <9h) - so no siamese for me so far.
", "rawMarkdown": "I am actually struggling with some of the bigger pretrained models (getting them to complete with reasonable param settings in <9h) - so no siamese for me so far.", "votes": 1, "replies": [ { "id": 512014, "postDate": "2019-04-10T14:02:09.123Z", "content": "correct me if I am wrong, but you can train locally and upload your model for inference ?
", "rawMarkdown": "correct me if I am wrong, but you can train locally and upload your model for inference ?" }, { "id": 512021, "postDate": "2019-04-10T14:07:26.800Z", "content": "I think it's legal
", "rawMarkdown": "I think it's legal" } ] }, { "id": 516324, "postDate": "2019-04-14T02:18:58.923Z", "content": "Maybe it works.
", "rawMarkdown": "Maybe it works." }, { "id": 509498, "postDate": "2019-04-07T21:45:08.803Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 518552, "author_name": "seefun", "author_url": "", "post_date": "2019-04-17T12:50:16.963000", "content": "Each picture has more than one labels. So I don't think triplet loss or any other metric learning method works well in this competition without some tricks like label clustering.
", "votes": 3, "replies": [ { "id": 518599, "author_name": "jionie", "author_url": "", "post_date": "2019-04-17T13:57:34.187000", "content": "Hmmmm, may I ask why do you think triplet loss won't work well if we can one hot encoded the labels?
", "votes": 2, "replies": [] }, { "id": 518601, "author_name": "seefun", "author_url": "", "post_date": "2019-04-17T14:00:24.057000", "content": "It 's just my guess. I think it should be verified in experiment.
", "votes": 3, "replies": [] }, { "id": 518624, "author_name": "", "author_url": "", "post_date": "2019-04-17T14:31:47.993000", "content": "I think it’s a good idea to encode the labels, let’s try it
", "votes": 1, "replies": [] }, { "id": 520475, "author_name": "jionie", "author_url": "", "post_date": "2019-04-21T04:24:23.810000", "content": "Though I haven't try, some loss functions like triplet loss + label clustering would be good. I think we cannot use triplet loss + one-hot encoding directly, some sub-classes should be correlated like (American) and (American or European). One-hot encoding would consider them as independent.
", "votes": 1, "replies": [] } ] }, { "id": 513205, "author_name": "xutao", "author_url": "", "post_date": "2019-04-11T03:28:57.897000", "content": "May not , Siamese network is a similarity measure method, which can be used to identify and classify categories when there are many categories but the number of samples per category is small. And the similarity of iMet dataset is not high as humpback whale identification.
", "votes": 3, "replies": [ { "id": 513232, "author_name": "", "author_url": "", "post_date": "2019-04-11T03:43:31.380000", "content": "Thank you
", "votes": 0, "replies": [] } ] }, { "id": 516334, "author_name": "interneuron", "author_url": "", "post_date": "2019-04-14T02:45:00.657000", "content": "Siamese network is quite powerful, I don’t see why it would not be successful here. But as others mentioned, it takes a very long time to train. Also, using the mpiotte Siamese net that was popular in whales will be tricky to use here without modifications, there are too many images to use exact lap solutions as was done there. You could try approximate lap. Not sure yet but classification seems like the way to start here.
", "votes": 1, "replies": [] }, { "id": 509345, "author_name": "Konrad Banachewicz", "author_url": "", "post_date": "2019-04-07T17:30:26.967000", "content": "I am actually struggling with some of the bigger pretrained models (getting them to complete with reasonable param settings in <9h) - so no siamese for me so far.
", "votes": 1, "replies": [ { "id": 512014, "author_name": "DrHB", "author_url": "", "post_date": "2019-04-10T14:02:09.123000", "content": "correct me if I am wrong, but you can train locally and upload your model for inference ?
", "votes": 0, "replies": [] }, { "id": 512021, "author_name": "", "author_url": "", "post_date": "2019-04-10T14:07:26.800000", "content": "I think it's legal
", "votes": 0, "replies": [] } ] }, { "id": 516324, "author_name": "", "author_url": "", "post_date": "2019-04-14T02:18:58.923000", "content": "Maybe it works.
", "votes": 0, "replies": [] }, { "id": 509498, "author_name": "", "author_url": "", "post_date": "2019-04-07T21:45:08.803000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "518552": "Each picture has more than one labels. So I don't think triplet loss or any other metric learning method works well in this competition without some tricks like label clustering.", "513205": "May not , Siamese network is a similarity measure method, which can be used to identify and classify categories when there are many categories but the number of samples per category is small. And the similarity of iMet dataset is not high as humpback whale identification.", "508498": "This challenge is a fine-grained image recognition task, so siamese is a popular solution.\nIn the previous competition [Humpback Whale Identification](https://www.kaggle.com/c/humpback-whale-identification), we used Siamese and it works well, I know the whale task is difference from iMet, graphs in Whale dataset are actually very similar, but iMet have much classes that each class is not similar.\nSiamese need to use more computing power and takes long time to train.\nso I want to ask if you have you tried to use siamese in this competition and if it works well?", "516334": "Siamese network is quite powerful, I don’t see why it would not be successful here. But as others mentioned, it takes a very long time to train. Also, using the mpiotte Siamese net that was popular in whales will be tricky to use here without modifications, there are too many images to use exact lap solutions as was done there. You could try approximate lap. Not sure yet but classification seems like the way to start here. ", "509345": "I am actually struggling with some of the bigger pretrained models (getting them to complete with reasonable param settings in <9h) - so no siamese for me so far.", "516324": "Maybe it works.", "509498": "" } }