I saw mixup works good in many image classification competition, so I tried to use it in this competition, but I'm wondering it is reasonable to apply it to regression task too or not.
\n\nActually my experiment was not successful, both for cv score and loss.\n- cv score without mixup: 0.9279874719\n- cv score with mixup: 0.9201798826
\n\nDoes anyone have idea, or achieve good result?\nThanks!
", "messages": [ { "id": "580629", "postDate": "07/20/2019 13:46:42", "content": "I saw mixup works good in many image classification competition, so I tried to use it in this competition, but I'm wondering it is reasonable to apply it to regression task too or not.
\n\nActually my experiment was not successful, both for cv score and loss.\n- cv score without mixup: 0.9279874719\n- cv score with mixup: 0.9201798826
\n\nDoes anyone have idea, or achieve good result?\nThanks!
", "rawMarkdown": "I saw mixup works good in many image classification competition, so I tried to use it in this competition, but I'm wondering it is reasonable to apply it to regression task too or not.\n\nActually my experiment was not successful, both for cv score and loss.\n- cv score without mixup: 0.9279874719\n- cv score with mixup: 0.9201798826\n\nDoes anyone have idea, or achieve good result?\nThanks!", "votes": null }, { "id": "580656", "postDate": "07/20/2019 14:19:28", "content": "I want to ask the same question
", "rawMarkdown": "I want to ask the same question", "votes": null }, { "id": "582044", "postDate": "07/22/2019 17:35:18", "content": "I think mixup makes sense for both classification as well as regression tasks but hasn't yet worked for me (coz the way you input and output is kinda same, wouldn't matter if categorical or not). That is though not the case with label smoothing which was often used along with classification tasks (seeing that our test data is also imbalanced), there needs to be a clever way of using it with regression!
", "rawMarkdown": "I think mixup makes sense for both classification as well as regression tasks but hasn't yet worked for me (coz the way you input and output is kinda same, wouldn't matter if categorical or not). That is though not the case with label smoothing which was often used along with classification tasks (seeing that our test data is also imbalanced), there needs to be a clever way of using it with regression!", "votes": null }, { "id": "582285", "postDate": "07/23/2019 02:39:11", "content": "Thank you for reply! I'll play around more and make a kernel for someone wants to try!
", "rawMarkdown": "Thank you for reply! I'll play around more and make a kernel for someone wants to try!", "votes": null }, { "id": "582674", "postDate": "07/23/2019 12:33:53", "content": "MIXUP parameters are difficult to adjust
", "rawMarkdown": "MIXUP parameters are difficult to adjust", "votes": null }, { "id": "582676", "postDate": "07/23/2019 12:34:17", "content": "MIXMATCH can try, but it may be spend more time
", "rawMarkdown": "MIXMATCH can try, but it may be spend more time", "votes": null }, { "id": "584808", "postDate": "07/26/2019 14:06:48", "content": "I implemented mixup, but it decreased both the CV and LB scores.\nLet me know if you have better results!
", "rawMarkdown": "I implemented mixup, but it decreased both the CV and LB scores.\nLet me know if you have better results!", "votes": null }, { "id": "585051", "postDate": "07/26/2019 21:09:21", "content": "Thanks, I've never tried to MixMatch, but it worth trying. If you have already tried, please share the result!
", "rawMarkdown": "Thanks, I've never tried to MixMatch, but it worth trying. If you have already tried, please share the result!", "votes": null }, { "id": "585052", "postDate": "07/26/2019 21:10:48", "content": "Yeah,,, me too. I think we should be careful with data augmentation and preprocess too.
", "rawMarkdown": "Yeah,,, me too. I think we should be careful with data augmentation and preprocess too.", "votes": null } ], "comments": [ { "id": 580656, "author_name": "xujingzhao", "author_url": "", "post_date": "07/20/2019 14:19:28", "content": "I want to ask the same question
", "votes": null, "replies": [] }, { "id": 582044, "author_name": "pranavmahajan725", "author_url": "", "post_date": "07/22/2019 17:35:18", "content": "I think mixup makes sense for both classification as well as regression tasks but hasn't yet worked for me (coz the way you input and output is kinda same, wouldn't matter if categorical or not). That is though not the case with label smoothing which was often used along with classification tasks (seeing that our test data is also imbalanced), there needs to be a clever way of using it with regression!
", "votes": null, "replies": [ { "id": 582285, "author_name": "bamps53", "author_url": "", "post_date": "07/23/2019 02:39:11", "content": "Thank you for reply! I'll play around more and make a kernel for someone wants to try!
", "votes": null, "replies": [] } ] }, { "id": 582674, "author_name": "icebergi", "author_url": "", "post_date": "07/23/2019 12:33:53", "content": "MIXUP parameters are difficult to adjust
", "votes": null, "replies": [] }, { "id": 582676, "author_name": "icebergi", "author_url": "", "post_date": "07/23/2019 12:34:17", "content": "MIXMATCH can try, but it may be spend more time
", "votes": null, "replies": [ { "id": 585051, "author_name": "bamps53", "author_url": "", "post_date": "07/26/2019 21:09:21", "content": "Thanks, I've never tried to MixMatch, but it worth trying. If you have already tried, please share the result!
", "votes": null, "replies": [] } ] }, { "id": 584808, "author_name": "nemethpeti", "author_url": "", "post_date": "07/26/2019 14:06:48", "content": "I implemented mixup, but it decreased both the CV and LB scores.\nLet me know if you have better results!
", "votes": null, "replies": [ { "id": 585052, "author_name": "bamps53", "author_url": "", "post_date": "07/26/2019 21:10:48", "content": "Yeah,,, me too. I think we should be careful with data augmentation and preprocess too.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "580629": "I saw mixup works good in many image classification competition, so I tried to use it in this competition, but I'm wondering it is reasonable to apply it to regression task too or not.\n\nActually my experiment was not successful, both for cv score and loss.\n- cv score without mixup: 0.9279874719\n- cv score with mixup: 0.9201798826\n\nDoes anyone have idea, or achieve good result?\nThanks!", "580656": "I want to ask the same question", "582044": "I think mixup makes sense for both classification as well as regression tasks but hasn't yet worked for me (coz the way you input and output is kinda same, wouldn't matter if categorical or not). That is though not the case with label smoothing which was often used along with classification tasks (seeing that our test data is also imbalanced), there needs to be a clever way of using it with regression!", "582285": "Thank you for reply! I'll play around more and make a kernel for someone wants to try!", "582674": "MIXUP parameters are difficult to adjust", "582676": "MIXMATCH can try, but it may be spend more time", "584808": "I implemented mixup, but it decreased both the CV and LB scores.\nLet me know if you have better results!", "585051": "Thanks, I've never tried to MixMatch, but it worth trying. If you have already tried, please share the result!", "585052": "Yeah,,, me too. I think we should be careful with data augmentation and preprocess too." }, "source": "meta" }