I've experimented with various CNN backbones, but they even don't perform well on the training data, and their training appears to be unstable (as seen in https://www.kaggle.com/code/sergeifironov/mobilenet-training/comments#2989784). Given their poor performance on the training data, are there any specific tricks to improve the results? Or are these models just not a good fit for this dataset?
", "messages": [ { "id": 3020131, "postDate": "2024-10-17T07:27:57.717Z", "content": "I believe complex modelling in general is not the solution for this competition. There's a discussion here on why linear models work best for this case and I mostly agree with it.
", "rawMarkdown": "I believe complex modelling in general is not the solution for this competition. There's a discussion here on why linear models work best for this case and I mostly agree with it.", "votes": 1 }, { "id": 3017817, "postDate": "2024-10-15T08:29:11.093Z", "content": "I've experimented with various CNN backbones, but they even don't perform well on the training data, and their training appears to be unstable (as seen in https://www.kaggle.com/code/sergeifironov/mobilenet-training/comments#2989784). Given their poor performance on the training data, are there any specific tricks to improve the results? Or are these models just not a good fit for this dataset?
", "rawMarkdown": "I've experimented with various CNN backbones, but they even don't perform well on the training data, and their training appears to be unstable (as seen in https://www.kaggle.com/code/sergeifironov/mobilenet-training/comments#2989784). Given their poor performance on the training data, are there any specific tricks to improve the results? Or are these models just not a good fit for this dataset?", "votes": 2 }, { "id": 3023522, "postDate": "2024-10-20T16:44:09.880Z", "content": "In theory, a CNN + resnet or transformer could cover linear models and archive better performance.
", "rawMarkdown": "In theory, a CNN + resnet or transformer could cover linear models and archive better performance." } ], "comments": [ { "id": 3020131, "author_name": "Natan Labarrère", "author_url": "", "post_date": "2024-10-17T07:27:57.717000", "content": "I believe complex modelling in general is not the solution for this competition. There's a discussion here on why linear models work best for this case and I mostly agree with it.
", "votes": 1, "replies": [] }, { "id": 3023522, "author_name": "Bobber Cheng", "author_url": "", "post_date": "2024-10-20T16:44:09.880000", "content": "In theory, a CNN + resnet or transformer could cover linear models and archive better performance.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3020131": "I believe complex modelling in general is not the solution for this competition. There's a discussion here on why linear models work best for this case and I mostly agree with it.", "3017817": "I've experimented with various CNN backbones, but they even don't perform well on the training data, and their training appears to be unstable (as seen in https://www.kaggle.com/code/sergeifironov/mobilenet-training/comments#2989784). Given their poor performance on the training data, are there any specific tricks to improve the results? Or are these models just not a good fit for this dataset?", "3023522": "In theory, a CNN + resnet or transformer could cover linear models and archive better performance." } }