Mixed Sample Data Augmentation (MSDA) has received increasing attention in recent years, with many successful variants such as MixUp and Cut- Mix. Following insight on the efficacy of Cut- Mix in particular, we propose FMix, an MSDA that uses binary masks obtained by applying a threshold to low frequency images sampled from Fourier space. FMix improves performance over MixUp and CutMix for a number of state-of-the- art models across a range of data sets and problem settings. We go on to analyse MixUp, CutMix, and FMix from an information theoretic perspect- ive, characterising learned models in terms of how they progressively compress the input with depth. Ultimately, our analyses allow us to decouple two complementary properties of augmentations, and present a unified framework for reasoning about MSDA.
\n\npaper：\nhttps://arxiv.xilesou.top/pdf/2002.12047.pdf
\n\ncode：\nhttps://github.com/ecs-vlc/FMix
", "messages": [ { "id": "761045", "postDate": "03/02/2020 04:23:22", "content": "Mixed Sample Data Augmentation (MSDA) has received increasing attention in recent years, with many successful variants such as MixUp and Cut- Mix. Following insight on the efficacy of Cut- Mix in particular, we propose FMix, an MSDA that uses binary masks obtained by applying a threshold to low frequency images sampled from Fourier space. FMix improves performance over MixUp and CutMix for a number of state-of-the- art models across a range of data sets and problem settings. We go on to analyse MixUp, CutMix, and FMix from an information theoretic perspect- ive, characterising learned models in terms of how they progressively compress the input with depth. Ultimately, our analyses allow us to decouple two complementary properties of augmentations, and present a unified framework for reasoning about MSDA.
\n\npaper：\nhttps://arxiv.xilesou.top/pdf/2002.12047.pdf
\n\ncode：\nhttps://github.com/ecs-vlc/FMix
", "rawMarkdown": "Mixed Sample Data Augmentation (MSDA) has received increasing attention in recent years, with many successful variants such as MixUp and Cut- Mix. Following insight on the efficacy of Cut- Mix in particular, we propose FMix, an MSDA that uses binary masks obtained by applying a threshold to low frequency images sampled from Fourier space. FMix improves performance over MixUp and CutMix for a number of state-of-the- art models across a range of data sets and problem settings. We go on to analyse MixUp, CutMix, and FMix from an information theoretic perspect- ive, characterising learned models in terms of how they progressively compress the input with depth. Ultimately, our analyses allow us to decouple two complementary properties of augmentations, and present a unified framework for reasoning about MSDA. \n\npaper：\nhttps://arxiv.xilesou.top/pdf/2002.12047.pdf\n\ncode：\nhttps://github.com/ecs-vlc/FMix", "votes": null }, { "id": "761067", "postDate": "03/02/2020 05:02:37", "content": "Very cool. Thanks for sharing.
", "rawMarkdown": "Very cool. Thanks for sharing.", "votes": null }, { "id": "761143", "postDate": "03/02/2020 07:24:54", "content": "hope it helps.
", "rawMarkdown": "hope it helps.", "votes": null }, { "id": "763187", "postDate": "03/04/2020 08:15:20", "content": "@machinelp did you try fmix for this competition? if yes then will you please tell us how it is performing for you? thank you
", "rawMarkdown": "machinelp did you try fmix for this competition? if yes then will you please tell us how it is performing for you? thank you", "votes": null }, { "id": "763196", "postDate": "03/04/2020 08:23:55", "content": "Sorry, I haven't used FMix in this competition yet.
", "rawMarkdown": "Sorry, I haven't used FMix in this competition yet.", "votes": null }, { "id": "763575", "postDate": "03/04/2020 16:05:33", "content": "I tried it. Didn't work as expected. But then I tried it in combination with mixup, cutmix. Maybe its like a lone wolf :D
", "rawMarkdown": "I tried it. Didn't work as expected. But then I tried it in combination with mixup, cutmix. Maybe its like a lone wolf :D", "votes": null }, { "id": "764596", "postDate": "03/05/2020 16:25:34", "content": "trained a few models with It, and I found that it's a lot more \"unstable\" than cutmix, causing loss to flucuate a lot more than with either cutmix or mixup or even both together. perhaps you can train witha much lower learning rate with this method but I havn't tried it.
", "rawMarkdown": "trained a few models with It, and I found that it's a lot more \"unstable\" than cutmix, causing loss to flucuate a lot more than with either cutmix or mixup or even both together. perhaps you can train witha much lower learning rate with this method but I havn't tried it.", "votes": null } ], "comments": [ { "id": 761067, "author_name": "cdeotte", "author_url": "", "post_date": "03/02/2020 05:02:37", "content": "Very cool. Thanks for sharing.
", "votes": null, "replies": [ { "id": 761143, "author_name": "machinelp", "author_url": "", "post_date": "03/02/2020 07:24:54", "content": "hope it helps.
", "votes": null, "replies": [] } ] }, { "id": 763187, "author_name": "mobassir", "author_url": "", "post_date": "03/04/2020 08:15:20", "content": "@machinelp did you try fmix for this competition? if yes then will you please tell us how it is performing for you? thank you
", "votes": null, "replies": [ { "id": 763196, "author_name": "machinelp", "author_url": "", "post_date": "03/04/2020 08:23:55", "content": "Sorry, I haven't used FMix in this competition yet.
", "votes": null, "replies": [] } ] }, { "id": 763575, "author_name": "mightyrains", "author_url": "", "post_date": "03/04/2020 16:05:33", "content": "I tried it. Didn't work as expected. But then I tried it in combination with mixup, cutmix. Maybe its like a lone wolf :D
", "votes": null, "replies": [] }, { "id": 764596, "author_name": "maxjon", "author_url": "", "post_date": "03/05/2020 16:25:34", "content": "trained a few models with It, and I found that it's a lot more \"unstable\" than cutmix, causing loss to flucuate a lot more than with either cutmix or mixup or even both together. perhaps you can train witha much lower learning rate with this method but I havn't tried it.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "761045": "Mixed Sample Data Augmentation (MSDA) has received increasing attention in recent years, with many successful variants such as MixUp and Cut- Mix. Following insight on the efficacy of Cut- Mix in particular, we propose FMix, an MSDA that uses binary masks obtained by applying a threshold to low frequency images sampled from Fourier space. FMix improves performance over MixUp and CutMix for a number of state-of-the- art models across a range of data sets and problem settings. We go on to analyse MixUp, CutMix, and FMix from an information theoretic perspect- ive, characterising learned models in terms of how they progressively compress the input with depth. Ultimately, our analyses allow us to decouple two complementary properties of augmentations, and present a unified framework for reasoning about MSDA. \n\npaper：\nhttps://arxiv.xilesou.top/pdf/2002.12047.pdf\n\ncode：\nhttps://github.com/ecs-vlc/FMix", "761067": "Very cool. Thanks for sharing.", "761143": "hope it helps.", "763187": "machinelp did you try fmix for this competition? if yes then will you please tell us how it is performing for you? thank you", "763196": "Sorry, I haven't used FMix in this competition yet.", "763575": "I tried it. Didn't work as expected. But then I tried it in combination with mixup, cutmix. Maybe its like a lone wolf :D", "764596": "trained a few models with It, and I found that it's a lot more \"unstable\" than cutmix, causing loss to flucuate a lot more than with either cutmix or mixup or even both together. perhaps you can train witha much lower learning rate with this method but I havn't tried it." }, "source": "meta" }