What has worked well so far for people, upsampling the cancer data or under sampling the 'no-cancer' data ?
", "messages": [ { "id": 2075838, "postDate": "2022-12-26T00:34:49.100Z", "content": "What has worked well so far for people, upsampling the cancer data or under sampling the 'no-cancer' data ?
", "rawMarkdown": "What has worked well so far for people, upsampling the cancer data or under sampling the 'no-cancer' data ?", "votes": 4 }, { "id": 2079554, "postDate": "2022-12-29T12:15:53.553Z", "content": "The majority of scientific papers which cover this topic suggest using oversampling over undersampling mostly due to the fact of removing data is not very productive.
\n\nAchieved way better results using oversampling instead of undersampling.
", "rawMarkdown": "The majority of scientific papers which cover this topic suggest using oversampling over undersampling mostly due to the fact of removing data is not very productive.\n\ne.g: https://www.researchgate.net/publication/340978368_Machine_Learning_with_Oversampling_and_Undersampling_Techniques_Overview_Study_and_Experimental_Results\n\nAchieved way better results using oversampling instead of undersampling.", "votes": 1 }, { "id": 2077286, "postDate": "2022-12-27T12:30:32.617Z", "content": "\"upsampling the cancer data\" == \"under sampling the no-cancer data\"
", "rawMarkdown": "\"upsampling the cancer data\" == \"under sampling the no-cancer data\"", "votes": 1 }, { "id": 2079942, "postDate": "2022-12-29T18:15:29.713Z", "content": "Neither of them worked for me in my previous experiences. I probably wasn't adjusting the training config into new sampled training set. If you can make them work, please share your secret :D
", "rawMarkdown": "Neither of them worked for me in my previous experiences. I probably wasn't adjusting the training config into new sampled training set. If you can make them work, please share your secret :D" }, { "id": 2077278, "postDate": "2022-12-27T12:15:33.117Z", "content": "Maybe I am wrong, but I've always thought if you uniformly under- or over-sample, then there is effectively no difference between them, except the undersampled dataset has fewer steps per epoch. In other words, if you were to train with the undersampled dataset len(oversampled)/len(undersampled) times as many epochs as the oversampled dataset, there should be no difference.
Oversampling the positive data points has worked better for me. Though, I managed to train the model with undersampling the negative data points. But I believe that using undersampling, it is harder to achieve superior model to that of oversampling.
", "rawMarkdown": "Oversampling the positive data points has worked better for me. Though, I managed to train the model with undersampling the negative data points. But I believe that using undersampling, it is harder to achieve superior model to that of oversampling.", "replies": [ { "id": 2076975, "postDate": "2022-12-27T04:33:58.517Z", "content": "Do you use the loss function with weight? Is this better than the upsample method? And I found that I was limited to LB 0.16. Do you have some suggestions? Thank you very much!
", "rawMarkdown": "Do you use the loss function with weight? Is this better than the upsample method? And I found that I was limited to LB 0.16. Do you have some suggestions? Thank you very much!\n", "replies": [ { "id": 2077183, "postDate": "2022-12-27T10:16:32.590Z", "content": "Yes, I use loss function with weight. Even with oversampling, you need to penalize wrong predictions on positive data points (cancer) more than that of negatives. So, you need a combination of both, oversampling and loss weighting. Also, try different models, not all models work equally in this competition according to my experience so far.
", "rawMarkdown": "Yes, I use loss function with weight. Even with oversampling, you need to penalize wrong predictions on positive data points (cancer) more than that of negatives. So, you need a combination of both, oversampling and loss weighting. Also, try different models, not all models work equally in this competition according to my experience so far.", "votes": 2 }, { "id": 2079862, "postDate": "2022-12-29T16:58:35.563Z", "content": "Thank you so much! I will try that.
", "rawMarkdown": "Thank you so much! I will try that.", "votes": 1 }, { "id": 2080157, "postDate": "2022-12-29T23:13:41.487Z", "content": "One more question.
\nI found that I only could get LB0.16.
\nI used efficientnetb4, weighted loss, balanced sampler, and augmentation. But I only could get 0.16 on LB best on 512*512 no ROI dataset. Do you have some suggestions? Do I miss something? ðŸ˜ðŸ˜
\nThank you so much.
", "rawMarkdown": "One more question.\n\nI found that I only could get LB0.16.\n\nI used efficientnetb4, weighted loss, balanced sampler, and augmentation. But I only could get 0.16 on LB best on 512*512 no ROI dataset. Do you have some suggestions? Do I miss something? ðŸ˜ðŸ˜\n\nThank you so much." }, { "id": 2080204, "postDate": "2022-12-30T00:19:44.720Z", "content": "You can try other models (like ResNext), higher resolutions (1024), Yolov5 based ROI extraction, training with auxiliary loss and label smoothing. We are all here to learn new things :)
", "rawMarkdown": "You can try other models (like ResNext), higher resolutions (1024), Yolov5 based ROI extraction, training with auxiliary loss and label smoothing. We are all here to learn new things :)", "votes": 2 }, { "id": 2080211, "postDate": "2022-12-30T00:26:54.573Z", "content": "Thank you very much! I will try that.
", "rawMarkdown": "Thank you very much! I will try that." } ] } ] } ], "comments": [ { "id": 2079554, "author_name": "Ali Abdin", "author_url": "", "post_date": "2022-12-29T12:15:53.553000", "content": "The majority of scientific papers which cover this topic suggest using oversampling over undersampling mostly due to the fact of removing data is not very productive.
\n\nAchieved way better results using oversampling instead of undersampling.
", "votes": 1, "replies": [] }, { "id": 2077286, "author_name": "Chenglu", "author_url": "", "post_date": "2022-12-27T12:30:32.617000", "content": "\"upsampling the cancer data\" == \"under sampling the no-cancer data\"
", "votes": 1, "replies": [] }, { "id": 2079942, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2022-12-29T18:15:29.713000", "content": "Neither of them worked for me in my previous experiences. I probably wasn't adjusting the training config into new sampled training set. If you can make them work, please share your secret :D
", "votes": 0, "replies": [] }, { "id": 2077278, "author_name": "James Howard", "author_url": "", "post_date": "2022-12-27T12:15:33.117000", "content": "Maybe I am wrong, but I've always thought if you uniformly under- or over-sample, then there is effectively no difference between them, except the undersampled dataset has fewer steps per epoch. In other words, if you were to train with the undersampled dataset len(oversampled)/len(undersampled) times as many epochs as the oversampled dataset, there should be no difference.
Oversampling the positive data points has worked better for me. Though, I managed to train the model with undersampling the negative data points. But I believe that using undersampling, it is harder to achieve superior model to that of oversampling.
", "votes": 0, "replies": [ { "id": 2076975, "author_name": "ChenxiangSun@NJU", "author_url": "", "post_date": "2022-12-27T04:33:58.517000", "content": "Do you use the loss function with weight? Is this better than the upsample method? And I found that I was limited to LB 0.16. Do you have some suggestions? Thank you very much!
", "votes": 0, "replies": [ { "id": 2077183, "author_name": "Rasoul Mojtahedzadeh", "author_url": "", "post_date": "2022-12-27T10:16:32.590000", "content": "Yes, I use loss function with weight. Even with oversampling, you need to penalize wrong predictions on positive data points (cancer) more than that of negatives. So, you need a combination of both, oversampling and loss weighting. Also, try different models, not all models work equally in this competition according to my experience so far.
", "votes": 2, "replies": [] }, { "id": 2079862, "author_name": "ChenxiangSun@NJU", "author_url": "", "post_date": "2022-12-29T16:58:35.563000", "content": "Thank you so much! I will try that.
", "votes": 1, "replies": [] }, { "id": 2080157, "author_name": "ChenxiangSun@NJU", "author_url": "", "post_date": "2022-12-29T23:13:41.487000", "content": "One more question.
\nI found that I only could get LB0.16.
\nI used efficientnetb4, weighted loss, balanced sampler, and augmentation. But I only could get 0.16 on LB best on 512*512 no ROI dataset. Do you have some suggestions? Do I miss something? ðŸ˜ðŸ˜
\nThank you so much.
", "votes": 0, "replies": [] }, { "id": 2080204, "author_name": "Rasoul Mojtahedzadeh", "author_url": "", "post_date": "2022-12-30T00:19:44.720000", "content": "You can try other models (like ResNext), higher resolutions (1024), Yolov5 based ROI extraction, training with auxiliary loss and label smoothing. We are all here to learn new things :)
", "votes": 2, "replies": [] }, { "id": 2080211, "author_name": "ChenxiangSun@NJU", "author_url": "", "post_date": "2022-12-30T00:26:54.573000", "content": "Thank you very much! I will try that.
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2075838": "What has worked well so far for people, upsampling the cancer data or under sampling the 'no-cancer' data ?", "2079554": "The majority of scientific papers which cover this topic suggest using oversampling over undersampling mostly due to the fact of removing data is not very productive.\n\ne.g: https://www.researchgate.net/publication/340978368_Machine_Learning_with_Oversampling_and_Undersampling_Techniques_Overview_Study_and_Experimental_Results\n\nAchieved way better results using oversampling instead of undersampling.", "2077286": "\"upsampling the cancer data\" == \"under sampling the no-cancer data\"", "2079942": "Neither of them worked for me in my previous experiences. I probably wasn't adjusting the training config into new sampled training set. If you can make them work, please share your secret :D", "2077278": "Maybe I am wrong, but I've always thought if you uniformly under- or over-sample, then there is effectively no difference between them, except the undersampled dataset has fewer steps per epoch. In other words, if you were to train with the undersampled dataset `len(oversampled)/len(undersampled)` times as many epochs as the oversampled dataset, there should be no difference.", "2076242": "Oversampling the positive data points has worked better for me. Though, I managed to train the model with undersampling the negative data points. But I believe that using undersampling, it is harder to achieve superior model to that of oversampling." } }