I think that Triplet loss(ReID) is a more effective approach than Classification in this problem.\nhttps://github.com/omoindrot/tensorflow-triplet-loss\nhttps://arxiv.org/abs/1503.03832\nhttp://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w6/Liao_Triplet-Based_Deep_Similarity_ICCV_2017_paper.pdf
\n\nHowever、that approach does not seem to work well as shown below\nhttps://www.kaggle.com/ateplyuk/keras-triplet-loss-lb-0-224\nhttps://www.kaggle.com/ashishpatel26/triplet-loss-network-for-humpback-whale-prediction
\n\nEven in my local, this approach quickly becomes over fitting. Is there anyone who tried a similar approach?
", "messages": [ { "id": 446687, "postDate": "2018-12-28T14:10:22.230Z", "content": "I think that Triplet loss(ReID) is a more effective approach than Classification in this problem.\nhttps://github.com/omoindrot/tensorflow-triplet-loss\nhttps://arxiv.org/abs/1503.03832\nhttp://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w6/Liao_Triplet-Based_Deep_Similarity_ICCV_2017_paper.pdf
\n\nHowever、that approach does not seem to work well as shown below\nhttps://www.kaggle.com/ateplyuk/keras-triplet-loss-lb-0-224\nhttps://www.kaggle.com/ashishpatel26/triplet-loss-network-for-humpback-whale-prediction
\n\nEven in my local, this approach quickly becomes over fitting. Is there anyone who tried a similar approach?
", "rawMarkdown": "I think that Triplet loss(ReID) is a more effective approach than Classification in this problem.\nhttps://github.com/omoindrot/tensorflow-triplet-loss\nhttps://arxiv.org/abs/1503.03832\nhttp://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w6/Liao_Triplet-Based_Deep_Similarity_ICCV_2017_paper.pdf\n\nHowever、that approach does not seem to work well as shown below\nhttps://www.kaggle.com/ateplyuk/keras-triplet-loss-lb-0-224\nhttps://www.kaggle.com/ashishpatel26/triplet-loss-network-for-humpback-whale-prediction\n\nEven in my local, this approach quickly becomes over fitting. Is there anyone who tried a similar approach?", "votes": 7 }, { "id": 449762, "postDate": "2019-01-03T17:46:57.183Z", "content": "I am investigating a triplet loss approach as well. The challenges I face currently is how to treat the 'new_whale' observations with this approach. I started to treat them just as negative examples but this looks limiting.
", "rawMarkdown": "I am investigating a triplet loss approach as well. The challenges I face currently is how to treat the 'new_whale' observations with this approach. I started to treat them just as negative examples but this looks limiting." }, { "id": 449281, "postDate": "2019-01-02T22:32:35.383Z", "rawMarkdown": "", "isDeleted": true }, { "id": 448936, "postDate": "2019-01-02T11:38:57.013Z", "content": "The triplet loss approach is pretty promising in our experiments (LB 0.78). The trick (for us) was to use the triplet loss in tensorflow.contrib. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/losses/python/metric_learning/metric_loss_ops.py
", "rawMarkdown": "The triplet loss approach is pretty promising in our experiments (LB 0.78). The trick (for us) was to use the triplet loss in tensorflow.contrib. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/losses/python/metric_learning/metric_loss_ops.py", "votes": 4, "isDeleted": true, "replies": [ { "id": 449132, "postDate": "2019-01-02T17:44:43.610Z", "content": "Thank you for your advice!!\nIn my trials, I used loss function in PyTorch. But it does not include semi hard negative mining trick. I think that it may be very important. I'm going to give it a try!! Thank you!!\nhttps://pytorch.org/docs/stable/nn.html#torch.nn.TripletMarginLoss
", "rawMarkdown": "Thank you for your advice!!\nIn my trials, I used loss function in PyTorch. But it does not include semi hard negative mining trick. I think that it may be very important. I'm going to give it a try!! Thank you!!\nhttps://pytorch.org/docs/stable/nn.html#torch.nn.TripletMarginLoss", "votes": 1 }, { "id": 449275, "postDate": "2019-01-02T22:24:27.240Z", "content": "I'm new to metric learning, and I'd like to adopt it as my approach. What is the difference between \"Triplet Loss\" and \"Triplet Semi-Hard Loss\"? I get the basic idea of Triplet Loss, in that we are trying to push negative examples further away from positive examples by a small margin (like a SVM). So, if I am understanding this correctly, a semi-hard triplet loss is using a semi-hard negative instead of a hard negative for training? Semi-hard being a negative that is not closer to the anchor compared to the positive, but still has a non-zero loss? Thanks, @Gert-Jan and @hirune924.
", "rawMarkdown": "I'm new to metric learning, and I'd like to adopt it as my approach. What is the difference between \"Triplet Loss\" and \"Triplet Semi-Hard Loss\"? I get the basic idea of Triplet Loss, in that we are trying to push negative examples further away from positive examples by a small margin (like a SVM). So, if I am understanding this correctly, a semi-hard triplet loss is using a semi-hard negative instead of a hard negative for training? Semi-hard being a negative that is not closer to the anchor compared to the positive, but still has a non-zero loss? Thanks, @Gert-Jan and @hirune924." }, { "id": 449282, "postDate": "2019-01-02T22:33:14.793Z", "content": "@Simeon Trieu:\nthere is a good explanation here by a French guy (in English): https://omoindrot.github.io/triplet-loss
\n\nEdit: actually this is the first link in the top msg... sorry
", "rawMarkdown": "@Simeon Trieu:\nthere is a good explanation here by a French guy (in English): https://omoindrot.github.io/triplet-loss\n\nEdit: actually this is the first link in the top msg... sorry", "votes": 1 }, { "id": 452786, "postDate": "2019-01-09T06:46:28.907Z", "content": "In your implementation, did you use detection to crop the image to get the score?
", "rawMarkdown": "In your implementation, did you use detection to crop the image to get the score?" }, { "id": 452836, "postDate": "2019-01-09T08:15:11.807Z", "content": "I have not used it yet. But I will try it.
", "rawMarkdown": "I have not used it yet. But I will try it." } ] } ], "comments": [ { "id": 449762, "author_name": "Tomasz Bartczak", "author_url": "", "post_date": "2019-01-03T17:46:57.183000", "content": "I am investigating a triplet loss approach as well. The challenges I face currently is how to treat the 'new_whale' observations with this approach. I started to treat them just as negative examples but this looks limiting.
", "votes": 0, "replies": [] }, { "id": 449281, "author_name": "", "author_url": "", "post_date": "2019-01-02T22:32:35.383000", "content": "", "votes": 0, "replies": [] }, { "id": 448936, "author_name": "", "author_url": "", "post_date": "2019-01-02T11:38:57.013000", "content": "The triplet loss approach is pretty promising in our experiments (LB 0.78). The trick (for us) was to use the triplet loss in tensorflow.contrib. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/losses/python/metric_learning/metric_loss_ops.py
", "votes": 4, "replies": [ { "id": 449132, "author_name": "hirune924", "author_url": "", "post_date": "2019-01-02T17:44:43.610000", "content": "Thank you for your advice!!\nIn my trials, I used loss function in PyTorch. But it does not include semi hard negative mining trick. I think that it may be very important. I'm going to give it a try!! Thank you!!\nhttps://pytorch.org/docs/stable/nn.html#torch.nn.TripletMarginLoss
", "votes": 1, "replies": [] }, { "id": 449275, "author_name": "Simeon Trieu", "author_url": "", "post_date": "2019-01-02T22:24:27.240000", "content": "I'm new to metric learning, and I'd like to adopt it as my approach. What is the difference between \"Triplet Loss\" and \"Triplet Semi-Hard Loss\"? I get the basic idea of Triplet Loss, in that we are trying to push negative examples further away from positive examples by a small margin (like a SVM). So, if I am understanding this correctly, a semi-hard triplet loss is using a semi-hard negative instead of a hard negative for training? Semi-hard being a negative that is not closer to the anchor compared to the positive, but still has a non-zero loss? Thanks, @Gert-Jan and @hirune924.
", "votes": 0, "replies": [] }, { "id": 449282, "author_name": "eagle4", "author_url": "", "post_date": "2019-01-02T22:33:14.793000", "content": "@Simeon Trieu:\nthere is a good explanation here by a French guy (in English): https://omoindrot.github.io/triplet-loss
\n\nEdit: actually this is the first link in the top msg... sorry
", "votes": 1, "replies": [] }, { "id": 452786, "author_name": "evan", "author_url": "", "post_date": "2019-01-09T06:46:28.907000", "content": "In your implementation, did you use detection to crop the image to get the score?
", "votes": 0, "replies": [] }, { "id": 452836, "author_name": "hirune924", "author_url": "", "post_date": "2019-01-09T08:15:11.807000", "content": "I have not used it yet. But I will try it.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "446687": "I think that Triplet loss(ReID) is a more effective approach than Classification in this problem.\nhttps://github.com/omoindrot/tensorflow-triplet-loss\nhttps://arxiv.org/abs/1503.03832\nhttp://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w6/Liao_Triplet-Based_Deep_Similarity_ICCV_2017_paper.pdf\n\nHowever、that approach does not seem to work well as shown below\nhttps://www.kaggle.com/ateplyuk/keras-triplet-loss-lb-0-224\nhttps://www.kaggle.com/ashishpatel26/triplet-loss-network-for-humpback-whale-prediction\n\nEven in my local, this approach quickly becomes over fitting. Is there anyone who tried a similar approach?", "449762": "I am investigating a triplet loss approach as well. The challenges I face currently is how to treat the 'new_whale' observations with this approach. I started to treat them just as negative examples but this looks limiting.", "449281": "", "448936": "The triplet loss approach is pretty promising in our experiments (LB 0.78). The trick (for us) was to use the triplet loss in tensorflow.contrib. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/losses/python/metric_learning/metric_loss_ops.py" } }