Since I plan on repeating the top solutions for this competition in a couple of weeks, I am spending some time reading some of the papers and tools shared by others. In that process I came across this interesting video and thought people may want to watch so here we go. It is from December 2018 London Machine Learning Meetup and is 46 minutes long but worth it imho.
\n\nhttps://www.youtube.com/watch?v=NpSpHlHpz6k
\n\nThe paper - On First-Order Meta-Learning Algorithms - https://arxiv.org/pdf/1803.02999.pdf
\n\nI will also add it to the knowledge round up list I have posted but thought people who have already read my post may not see it.
", "messages": [ { "id": "483191", "postDate": "03/04/2019 09:27:41", "content": "Since I plan on repeating the top solutions for this competition in a couple of weeks, I am spending some time reading some of the papers and tools shared by others. In that process I came across this interesting video and thought people may want to watch so here we go. It is from December 2018 London Machine Learning Meetup and is 46 minutes long but worth it imho.
\n\nhttps://www.youtube.com/watch?v=NpSpHlHpz6k
\n\nThe paper - On First-Order Meta-Learning Algorithms - https://arxiv.org/pdf/1803.02999.pdf
\n\nI will also add it to the knowledge round up list I have posted but thought people who have already read my post may not see it.
", "rawMarkdown": "Since I plan on repeating the top solutions for this competition in a couple of weeks, I am spending some time reading some of the [papers and tools](https://www.kaggle.com/c/humpback-whale-identification/discussion/82437) shared by others. In that process I came across this interesting video and thought people may want to watch so here we go. It is from December 2018 London Machine Learning Meetup and is 46 minutes long but worth it imho.\n\nhttps://www.youtube.com/watch?v=NpSpHlHpz6k\n\nThe paper - On First-Order Meta-Learning Algorithms - https://arxiv.org/pdf/1803.02999.pdf\n\nI will also add it to the knowledge round up list I have posted but thought people who have already read [my post](https://www.kaggle.com/c/humpback-whale-identification/discussion/82437) may not see it.", "votes": null }, { "id": "483297", "postDate": "03/04/2019 12:55:48", "content": "I'm also planning to repeat some of the shared solution to see the influence of each method proposed method.
\n\nBecause in general we can see trend that:\n1. Triplet + SoftMax (1st place)\n2. ArcFace\n3. Siamese
\n\nThey general woks pretty much the same, but maybe merging them can boost score a little bit also :)
\n\nAlso I'm interested for best model without any TTA, post-processing etc.
", "rawMarkdown": "I'm also planning to repeat some of the shared solution to see the influence of each method proposed method.\n\nBecause in general we can see trend that:\n1. Triplet + SoftMax (1st place)\n2. ArcFace\n3. Siamese\n\nThey general woks pretty much the same, but maybe merging them can boost score a little bit also :)\n\nAlso I'm interested for best model without any TTA, post-processing etc.", "votes": null }, { "id": "483362", "postDate": "03/04/2019 14:04:27", "content": "I'm also planning to repeat some of the top solution.\nand will you release your code?\nThanks
", "rawMarkdown": "I'm also planning to repeat some of the top solution.\nand will you release your code?\nThanks", "votes": null }, { "id": "483554", "postDate": "03/04/2019 19:56:13", "content": "@melgor and @excllent123, this competition being a kind of an all inclusive problem, I figured some of us will not be done with it yet. Like Heng said it includes classification, metric learning, few shot learning (KNN classifier), image retrieval and distractor/outlier rejection (aka one-class training). That is what makes it special :-)
", "rawMarkdown": "melgor and @excllent123, this competition being a kind of an all inclusive problem, I figured some of us will not be done with it yet. Like Heng said it includes classification, metric learning, few shot learning (KNN classifier), image retrieval and distractor/outlier rejection (aka one-class training). That is what makes it special :-)", "votes": null } ], "comments": [ { "id": 483297, "author_name": "melgor", "author_url": "", "post_date": "03/04/2019 12:55:48", "content": "I'm also planning to repeat some of the shared solution to see the influence of each method proposed method.
\n\nBecause in general we can see trend that:\n1. Triplet + SoftMax (1st place)\n2. ArcFace\n3. Siamese
\n\nThey general woks pretty much the same, but maybe merging them can boost score a little bit also :)
\n\nAlso I'm interested for best model without any TTA, post-processing etc.
", "votes": null, "replies": [ { "id": 483362, "author_name": "excllent123", "author_url": "", "post_date": "03/04/2019 14:04:27", "content": "I'm also planning to repeat some of the top solution.\nand will you release your code?\nThanks
", "votes": null, "replies": [] }, { "id": 483554, "author_name": "sheriytm", "author_url": "", "post_date": "03/04/2019 19:56:13", "content": "@melgor and @excllent123, this competition being a kind of an all inclusive problem, I figured some of us will not be done with it yet. Like Heng said it includes classification, metric learning, few shot learning (KNN classifier), image retrieval and distractor/outlier rejection (aka one-class training). That is what makes it special :-)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "483191": "Since I plan on repeating the top solutions for this competition in a couple of weeks, I am spending some time reading some of the [papers and tools](https://www.kaggle.com/c/humpback-whale-identification/discussion/82437) shared by others. In that process I came across this interesting video and thought people may want to watch so here we go. It is from December 2018 London Machine Learning Meetup and is 46 minutes long but worth it imho.\n\nhttps://www.youtube.com/watch?v=NpSpHlHpz6k\n\nThe paper - On First-Order Meta-Learning Algorithms - https://arxiv.org/pdf/1803.02999.pdf\n\nI will also add it to the knowledge round up list I have posted but thought people who have already read [my post](https://www.kaggle.com/c/humpback-whale-identification/discussion/82437) may not see it.", "483297": "I'm also planning to repeat some of the shared solution to see the influence of each method proposed method.\n\nBecause in general we can see trend that:\n1. Triplet + SoftMax (1st place)\n2. ArcFace\n3. Siamese\n\nThey general woks pretty much the same, but maybe merging them can boost score a little bit also :)\n\nAlso I'm interested for best model without any TTA, post-processing etc.", "483362": "I'm also planning to repeat some of the top solution.\nand will you release your code?\nThanks", "483554": "melgor and @excllent123, this competition being a kind of an all inclusive problem, I figured some of us will not be done with it yet. Like Heng said it includes classification, metric learning, few shot learning (KNN classifier), image retrieval and distractor/outlier rejection (aka one-class training). That is what makes it special :-)" }, "source": "meta" }