![\"Snipaste-2022-02-19-12-00-27.png\"](\"https://i.postimg.cc/MHRdxgFF/Snipaste-2022-02-19-12-00-27.png\")
Inspired by this link. I was reminded of a hierarchical multi-label task I had done before. Just like the data provided in this competition, we can divide the labels of each image into primary labels and secondary labels, and these two labels are related to each other, such tasks are called hierarchical multi-label tasks (HMC). If you want to train a classifier, but worry about having only one image for many individuals, this model might help you. It's just a classification header, maybe it can help you get better features.
\n In this paper, a novel neural network architectures for HMC called HMCN is proposed, capable of simultaneously optimizing local and global loss functions for discovering local hierarchical class-relationships and global information from the entire class hierarchy while penalizing hierarchical violations.
\n This model has good performance in my previous work, but I'm new to the competition and haven't tried it here, but I hope this helps. The schematic diagram of the model is as follows，
\n[url=https://postimg.cc/GHKP7yCh][img]https://i.postimg.cc/GHKP7yCh/Snipaste-2022-02-19-12-00-27.png[/img][/url]
\n Hope to help everyone！
Inspired by this link. I was reminded of a hierarchical multi-label task I had done before. Just like the data provided in this competition, we can divide the labels of each image into primary labels and secondary labels, and these two labels are related to each other, such tasks are called hierarchical multi-label tasks (HMC). If you want to train a classifier, but worry about having only one image for many individuals, this model might help you. It's just a classification header, maybe it can help you get better features.
\n In this paper, a novel neural network architectures for HMC called HMCN is proposed, capable of simultaneously optimizing local and global loss functions for discovering local hierarchical class-relationships and global information from the entire class hierarchy while penalizing hierarchical violations.
\n This model has good performance in my previous work, but I'm new to the competition and haven't tried it here, but I hope this helps. The schematic diagram of the model is as follows，
\n[url=https://postimg.cc/GHKP7yCh][img]https://i.postimg.cc/GHKP7yCh/Snipaste-2022-02-19-12-00-27.png[/img][/url]
\n Hope to help everyone！
After a month, I finally have time to do this experiment, I just used it briefly, it seems to have an effect on my lb, my lb score is from 0.774 -> 0.776. This gives me hope, so I would like to carry out further exploration, hoping to achieve results.✌️
", "rawMarkdown": "After a month, I finally have time to do this experiment, I just used it briefly, it seems to have an effect on my lb, my lb score is from 0.774 -> 0.776. This gives me hope, so I would like to carry out further exploration, hoping to achieve results.✌️", "votes": null } ], "comments": [ { "id": 1696743, "author_name": "zhuwanglju", "author_url": "", "post_date": "02/19/2022 04:34:10", "content": "
After a month, I finally have time to do this experiment, I just used it briefly, it seems to have an effect on my lb, my lb score is from 0.774 -> 0.776. This gives me hope, so I would like to carry out further exploration, hoping to achieve results.✌️
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1696742": "Inspired by this [link](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/307286). I was reminded of a hierarchical multi-label task I had done before. Just like the data provided in this competition, we can divide the labels of each image into primary labels and secondary labels, and these two labels are related to each other, such tasks are called hierarchical multi-label tasks (HMC). If you want to train a classifier, but worry about having only one image for many individuals, this model might help you. It's just a classification header, maybe it can help you get better features.\n\n\n In this [paper](https://proceedings.mlr.press/v80/wehrmann18a/wehrmann18a.pdf), a novel neural network architectures for HMC called HMCN is proposed, capable of simultaneously optimizing local and global loss functions for discovering local hierarchical class-relationships and global information from the entire class hierarchy while penalizing hierarchical violations.\n\n\n This model has good performance in my previous work, but I'm new to the competition and haven't tried it here, but I hope this helps. The schematic diagram of the model is as follows，\n[url=https://postimg.cc/GHKP7yCh][img]https://i.postimg.cc/GHKP7yCh/Snipaste-2022-02-19-12-00-27.png[/img][/url]\n\n\n Hope to help everyone！", "1696743": "[](https://postimg.cc/rRVSNZc1)", "1743752": "After a month, I finally have time to do this experiment, I just used it briefly, it seems to have an effect on my lb, my lb score is from 0.774 -> 0.776. This gives me hope, so I would like to carry out further exploration, hoping to achieve results.✌️" }, "source": "meta" }