![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1977801%2Fd0138e5d3169deeaee7ecf80404c33cf%2FBengalai_head.png?generation=1584608889278316&alt=media\")
I'm very surprised at the big shake and I don't still understand why I got this place.
\n\nI suppose one of reasons why I got 10th place by very simple solution is using RandomErasing but not using CutMix/MixUp.
\n\nI think these two are effective in making unseen combinations of components, but not effective in disentangling interdependence of components in seen graphemes.
\n\nWhat do you think about this?
", "messages": [ { "id": "778955", "postDate": "03/18/2020 22:09:39", "content": "I'm very surprised at the big shake and I don't still understand why I got this place.
\n\nI suppose one of reasons why I got 10th place by very simple solution is using RandomErasing but not using CutMix/MixUp.
\n\nI think these two are effective in making unseen combinations of components, but not effective in disentangling interdependence of components in seen graphemes.
\n\nWhat do you think about this?
", "rawMarkdown": "I'm very surprised at the big shake and I don't still understand why I got this place.\n\nI suppose one of reasons why [I got 10th place by very simple solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136815) is using RandomErasing **but not using CutMix/MixUp**.\n\nI think these two are effective in making unseen combinations of components, but **not effective in disentangling interdependence of components in seen graphemes.**\n\nWhat do you think about this?", "votes": null }, { "id": "779153", "postDate": "03/19/2020 04:13:51", "content": "Late sub\n||pucblic|private|\n|---| --- | --- |\n|cutmix|0.9876 |0.9246|\n|no cutmix|0.9830|0.9201|
\n\nThe score goes down without cutmix in my model.😓
", "rawMarkdown": "Late sub\n||pucblic|private|\n|---| --- | --- |\n|cutmix|0.9876 |0.9246|\n|no cutmix|0.9830|0.9201|\n\nThe score goes down without cutmix in my model.😓", "votes": null }, { "id": "779260", "postDate": "03/19/2020 06:42:55", "content": "Thanks for sharing your result!
\n\nDid you use a simple mask augmentation such as CutOut/RandomErasing?
", "rawMarkdown": "Thanks for sharing your result!\n\nDid you use a simple mask augmentation such as CutOut/RandomErasing?", "votes": null }, { "id": "779272", "postDate": "03/19/2020 06:56:03", "content": "Yes, I use shiftscalerotate and CutOut. \nMaybe, My model only classify seen grapheme and cutmix can be ​​effective for seen grapheme.
", "rawMarkdown": "Yes, I use shiftscalerotate and CutOut. \nMaybe, My model only classify seen grapheme and cutmix can be ​​effective for seen grapheme.", "votes": null }, { "id": "779288", "postDate": "03/19/2020 07:06:23", "content": "This is quite interesting results. Thanks for sharing it!\nMaybe the secret sause appears when we combine random erasing and se average pooling block..?
", "rawMarkdown": "This is quite interesting results. Thanks for sharing it!\nMaybe the secret sause appears when we combine random erasing and se average pooling block..?", "votes": null }, { "id": "779316", "postDate": "03/19/2020 07:48:29", "content": "Hmm..., it is interesting.\nAs @kyoshioka47 says, it is important to use not only RandomErasing but component-wise spatial weighted average pooling like sSE-Pooling ?
\n\n@d1348k , how about a pooling layer in your model ?
", "rawMarkdown": "Hmm..., it is interesting.\nAs @kyoshioka47 says, it is important to use not only RandomErasing but component-wise spatial weighted average pooling like sSE-Pooling ?\n\n@d1348k , how about a pooling layer in your model ?", "votes": null }, { "id": "779320", "postDate": "03/19/2020 07:49:01", "content": "I did raise some questions about the use of CutMix. While I do think it does help, I also think it has to be applied more thoughtfully than at random. People have been sharing CAM-CutMix which IMHO is a really interesting approach to tackle this and not do it purely at random.
", "rawMarkdown": "I did raise [some questions about the use of CutMix](https://www.kaggle.com/c/bengaliai-cv19/discussion/126557). While I do think it does help, I also think it has to be applied more thoughtfully than at random. People have been sharing CAM-CutMix which IMHO is a really interesting approach to tackle this and not do it purely at random.", "votes": null }, { "id": "779366", "postDate": "03/19/2020 09:09:51", "content": "My model use GeM. Here is my head.
\n\n
Thanks. Your model seems to have component-wise necks before pooling layer 🤔 Hmm...
", "rawMarkdown": "Thanks. Your model seems to have component-wise necks before pooling layer 🤔 Hmm...", "votes": null }, { "id": "779563", "postDate": "03/19/2020 13:30:04", "content": "Thanks! That post is helpful for me to understand the risk of using CutMix roughly.
", "rawMarkdown": "Thanks! That post is helpful for me to understand the risk of using CutMix roughly.", "votes": null } ], "comments": [ { "id": 779153, "author_name": "d1348k", "author_url": "", "post_date": "03/19/2020 04:13:51", "content": "Late sub\n||pucblic|private|\n|---| --- | --- |\n|cutmix|0.9876 |0.9246|\n|no cutmix|0.9830|0.9201|
\n\nThe score goes down without cutmix in my model.😓
", "votes": null, "replies": [ { "id": 779260, "author_name": "ttahara", "author_url": "", "post_date": "03/19/2020 06:42:55", "content": "Thanks for sharing your result!
\n\nDid you use a simple mask augmentation such as CutOut/RandomErasing?
", "votes": null, "replies": [] }, { "id": 779272, "author_name": "d1348k", "author_url": "", "post_date": "03/19/2020 06:56:03", "content": "Yes, I use shiftscalerotate and CutOut. \nMaybe, My model only classify seen grapheme and cutmix can be ​​effective for seen grapheme.
", "votes": null, "replies": [] }, { "id": 779288, "author_name": "kyoshioka47", "author_url": "", "post_date": "03/19/2020 07:06:23", "content": "This is quite interesting results. Thanks for sharing it!\nMaybe the secret sause appears when we combine random erasing and se average pooling block..?
", "votes": null, "replies": [] }, { "id": 779316, "author_name": "ttahara", "author_url": "", "post_date": "03/19/2020 07:48:29", "content": "Hmm..., it is interesting.\nAs @kyoshioka47 says, it is important to use not only RandomErasing but component-wise spatial weighted average pooling like sSE-Pooling ?
\n\n@d1348k , how about a pooling layer in your model ?
", "votes": null, "replies": [] }, { "id": 779366, "author_name": "d1348k", "author_url": "", "post_date": "03/19/2020 09:09:51", "content": "My model use GeM. Here is my head.
\n\n
Thanks. Your model seems to have component-wise necks before pooling layer 🤔 Hmm...
", "votes": null, "replies": [] } ] }, { "id": 779320, "author_name": "maxlenormand", "author_url": "", "post_date": "03/19/2020 07:49:01", "content": "I did raise some questions about the use of CutMix. While I do think it does help, I also think it has to be applied more thoughtfully than at random. People have been sharing CAM-CutMix which IMHO is a really interesting approach to tackle this and not do it purely at random.
", "votes": null, "replies": [ { "id": 779563, "author_name": "ttahara", "author_url": "", "post_date": "03/19/2020 13:30:04", "content": "Thanks! That post is helpful for me to understand the risk of using CutMix roughly.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "778955": "I'm very surprised at the big shake and I don't still understand why I got this place.\n\nI suppose one of reasons why [I got 10th place by very simple solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136815) is using RandomErasing **but not using CutMix/MixUp**.\n\nI think these two are effective in making unseen combinations of components, but **not effective in disentangling interdependence of components in seen graphemes.**\n\nWhat do you think about this?", "779153": "Late sub\n||pucblic|private|\n|---| --- | --- |\n|cutmix|0.9876 |0.9246|\n|no cutmix|0.9830|0.9201|\n\nThe score goes down without cutmix in my model.😓", "779260": "Thanks for sharing your result!\n\nDid you use a simple mask augmentation such as CutOut/RandomErasing?", "779272": "Yes, I use shiftscalerotate and CutOut. \nMaybe, My model only classify seen grapheme and cutmix can be ​​effective for seen grapheme.", "779288": "This is quite interesting results. Thanks for sharing it!\nMaybe the secret sause appears when we combine random erasing and se average pooling block..?", "779316": "Hmm..., it is interesting.\nAs @kyoshioka47 says, it is important to use not only RandomErasing but component-wise spatial weighted average pooling like sSE-Pooling ?\n\n@d1348k , how about a pooling layer in your model ?", "779320": "I did raise [some questions about the use of CutMix](https://www.kaggle.com/c/bengaliai-cv19/discussion/126557). While I do think it does help, I also think it has to be applied more thoughtfully than at random. People have been sharing CAM-CutMix which IMHO is a really interesting approach to tackle this and not do it purely at random.", "779366": "My model use GeM. Here is my head.\n\n", "779555": "Thanks. Your model seems to have component-wise necks before pooling layer 🤔 Hmm...", "779563": "Thanks! That post is helpful for me to understand the risk of using CutMix roughly." }, "source": "meta" }