Really happy I took part in this competition. My first one, so I learned a lot. Thanks to all including my teammate liovch (who also inspired this post) and those kind enough to share interesting kernels/ideas on the forum.
\n\nJust a thought: Wouldn't it be great if we could still submit and evaluate scores for a few days after the deadline has passed? Not to compete, but to assess the ideas we couldn't test so we can learn a bit more before it's out of our hearts and heads.
", "messages": [ { "id": "792566", "postDate": "03/31/2020 10:35:04", "content": "Really happy I took part in this competition. My first one, so I learned a lot. Thanks to all including my teammate liovch (who also inspired this post) and those kind enough to share interesting kernels/ideas on the forum.
\n\nJust a thought: Wouldn't it be great if we could still submit and evaluate scores for a few days after the deadline has passed? Not to compete, but to assess the ideas we couldn't test so we can learn a bit more before it's out of our hearts and heads.
", "rawMarkdown": "Really happy I took part in this competition. My first one, so I learned a lot. Thanks to all including my teammate liovch (who also inspired this post) and those kind enough to share interesting kernels/ideas on the forum.\n\nJust a thought: Wouldn't it be great if we could still submit and evaluate scores for a few days after the deadline has passed? Not to compete, but to assess the ideas we couldn't test so we can learn a bit more before it's out of our hearts and heads.", "votes": null }, { "id": "793277", "postDate": "03/31/2020 23:21:29", "content": "I will definitely do that! I joined the competition relatively late - last month. I wanted to implement unsupervised data augmentation for this challenge to try out some fancy custom augmentation ideas. Given the fact overfitting was the biggest challenge I really wanted to see what augmentation techniques would work best. Only had time to implement one kind of custom augmentation similar to cutmix. Let's see how we will end on private LB :)
", "rawMarkdown": "I will definitely do that! I joined the competition relatively late - last month. I wanted to implement unsupervised data augmentation for this challenge to try out some fancy custom augmentation ideas. Given the fact overfitting was the biggest challenge I really wanted to see what augmentation techniques would work best. Only had time to implement one kind of custom augmentation similar to cutmix. Let's see how we will end on private LB :)", "votes": null }, { "id": "793347", "postDate": "04/01/2020 00:51:30", "content": "Thats interesting. Any tips on how you implemented cutmix?
", "rawMarkdown": "Thats interesting. Any tips on how you implemented cutmix?", "votes": null }, { "id": "793848", "postDate": "04/01/2020 10:49:17", "content": "Sure. First, let me mention my model uses face crops for training. I don't know if it could be done better but basically, I horizontally joined 2 faces (50% of the original face crop and 50% of a random face crop from the same class). This gave a 0.02 boost on public LB. I chose a horizontal join of 50% since I assumed fake alteration is symmetric.
\n\nThe idea was to not let model overfit to specific faces/actors by memorizing them. So having this augmentation allowed to create more faces by joining halves of two different actors. This notebook shows the exact augmentation.
", "rawMarkdown": "Sure. First, let me mention my model uses face crops for training. I don't know if it could be done better but basically, I horizontally joined 2 faces (50% of the original face crop and 50% of a random face crop from the same class). This gave a 0.02 boost on public LB. I chose a horizontal join of 50% since I assumed fake alteration is symmetric.\n\nThe idea was to not let model overfit to specific faces/actors by memorizing them. So having this augmentation allowed to create more faces by joining halves of two different actors. This [notebook](https://github.com/KeremTurgutlu/dfdc/blob/master/nbs/23_single_frame_model_randmerge.ipynb) shows the exact augmentation.", "votes": null }, { "id": "793852", "postDate": "04/01/2020 10:54:53", "content": "Clever!
", "rawMarkdown": "Clever!", "votes": null } ], "comments": [ { "id": 793277, "author_name": "keremt", "author_url": "", "post_date": "03/31/2020 23:21:29", "content": "I will definitely do that! I joined the competition relatively late - last month. I wanted to implement unsupervised data augmentation for this challenge to try out some fancy custom augmentation ideas. Given the fact overfitting was the biggest challenge I really wanted to see what augmentation techniques would work best. Only had time to implement one kind of custom augmentation similar to cutmix. Let's see how we will end on private LB :)
", "votes": null, "replies": [ { "id": 793347, "author_name": "shinig4mi", "author_url": "", "post_date": "04/01/2020 00:51:30", "content": "Thats interesting. Any tips on how you implemented cutmix?
", "votes": null, "replies": [] }, { "id": 793848, "author_name": "keremt", "author_url": "", "post_date": "04/01/2020 10:49:17", "content": "Sure. First, let me mention my model uses face crops for training. I don't know if it could be done better but basically, I horizontally joined 2 faces (50% of the original face crop and 50% of a random face crop from the same class). This gave a 0.02 boost on public LB. I chose a horizontal join of 50% since I assumed fake alteration is symmetric.
\n\nThe idea was to not let model overfit to specific faces/actors by memorizing them. So having this augmentation allowed to create more faces by joining halves of two different actors. This notebook shows the exact augmentation.
", "votes": null, "replies": [] }, { "id": 793852, "author_name": "mpware", "author_url": "", "post_date": "04/01/2020 10:54:53", "content": "Clever!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "792566": "Really happy I took part in this competition. My first one, so I learned a lot. Thanks to all including my teammate liovch (who also inspired this post) and those kind enough to share interesting kernels/ideas on the forum.\n\nJust a thought: Wouldn't it be great if we could still submit and evaluate scores for a few days after the deadline has passed? Not to compete, but to assess the ideas we couldn't test so we can learn a bit more before it's out of our hearts and heads.", "793277": "I will definitely do that! I joined the competition relatively late - last month. I wanted to implement unsupervised data augmentation for this challenge to try out some fancy custom augmentation ideas. Given the fact overfitting was the biggest challenge I really wanted to see what augmentation techniques would work best. Only had time to implement one kind of custom augmentation similar to cutmix. Let's see how we will end on private LB :)", "793347": "Thats interesting. Any tips on how you implemented cutmix?", "793848": "Sure. First, let me mention my model uses face crops for training. I don't know if it could be done better but basically, I horizontally joined 2 faces (50% of the original face crop and 50% of a random face crop from the same class). This gave a 0.02 boost on public LB. I chose a horizontal join of 50% since I assumed fake alteration is symmetric.\n\nThe idea was to not let model overfit to specific faces/actors by memorizing them. So having this augmentation allowed to create more faces by joining halves of two different actors. This [notebook](https://github.com/KeremTurgutlu/dfdc/blob/master/nbs/23_single_frame_model_randmerge.ipynb) shows the exact augmentation.", "793852": "Clever!" }, "source": "meta" }