mix-based augmentations are reported helpful in many kaggle image competitions, cutmix, mixup, fmix, etc, but I didn't see any notebook using mix-based aug, is it because the competition is multi-label and hard to adapt label?
", "messages": [ { "id": "1220673", "postDate": "02/28/2021 08:41:25", "content": "mix-based augmentations are reported helpful in many kaggle image competitions, cutmix, mixup, fmix, etc, but I didn't see any notebook using mix-based aug, is it because the competition is multi-label and hard to adapt label?
", "rawMarkdown": "mix-based augmentations are reported helpful in many kaggle image competitions, cutmix, mixup, fmix, etc, but I didn't see any notebook using mix-based aug, is it because the competition is multi-label and hard to adapt label?", "votes": null }, { "id": "1220695", "postDate": "02/28/2021 09:00:45", "content": "Adapting CutMix or MixUp to multi-label problems isn't hard, e.g. for CutMix instead of one label getting mixed, we mix all as
\n$$w_i \\times y_{kj} + (1-w_i) \\times y_{ij}$$
\nfor the jth label of image i that is having a proportion w_i of its area replaced by something cut from image k. Standard implementations in some packages may not cover it, but it's easy enough to implement.
I assume some people have experimented with it, but for me it's not been in the top of my list. The slightly tricky bit is that e.g. for CutMix, if you copy a bit of an image that does not show the anything relevant to the catheter placement, it does not actually result in a mixing of the classification of the images as far as a human is concerned. On the other hand, if you copy a bit showing a wrong placement, this changes the classification totally, not partially. It's not like photos of cats and dogs, where if you see half a dog and half a cat, it's reasonable to say 0.5 for each class. Perhaps something like SnapMix would do that better?! In any case, that's my speculation and if it ends up working well, then perhaps I'm wrong about that.
", "rawMarkdown": "Adapting CutMix or MixUp to multi-label problems isn't hard, e.g. for CutMix instead of one label getting mixed, we mix all as \n$$w_i \\times y_{kj} + (1-w_i) \\times y_{ij}$$ \nfor the jth label of image i that is having a proportion w_i of its area replaced by something cut from image k. Standard implementations in some packages may not cover it, but it's easy enough to implement.\n\nI assume some people have experimented with it, but for me it's not been in the top of my list. The slightly tricky bit is that e.g. for CutMix, if you copy a bit of an image that does not show the anything relevant to the catheter placement, it does not actually result in a mixing of the classification of the images as far as a human is concerned. On the other hand, if you copy a bit showing a wrong placement, this changes the classification totally, not partially. It's not like photos of cats and dogs, where if you see half a dog and half a cat, it's reasonable to say 0.5 for each class. Perhaps something like SnapMix would do that better?! In any case, that's my speculation and if it ends up working well, then perhaps I'm wrong about that.", "votes": null }, { "id": "1220741", "postDate": "02/28/2021 10:07:35", "content": "Mix-based aug is usually useful but need relatively more computing resource. If we use strong mix-based aug with light transforms, the hyperparameter required to tune will reduced to few. So lazy boy like me often use these method.
", "rawMarkdown": "Mix-based aug is usually useful but need relatively more computing resource. If we use strong mix-based aug with light transforms, the hyperparameter required to tune will reduced to few. So lazy boy like me often use these method.", "votes": null }, { "id": "1220861", "postDate": "02/28/2021 12:25:18", "content": "I tested both Mixup and CutMix but they did not help improving performance (even they were far worse). Actually, the convergence speed was too slow. Maybe I should train with more epochs.
", "rawMarkdown": "I tested both Mixup and CutMix but they did not help improving performance (even they were far worse). Actually, the convergence speed was too slow. Maybe I should train with more epochs.", "votes": null }, { "id": "1221594", "postDate": "03/01/2021 05:41:22", "content": "In this competition dataset of lungs X-rays imgs are are taken in a fixed manner (means vertical) so here only HFlip and normalization will improve results. But you can also go with heavy augmentation it may increase your results a little bit on large models like Resnet200.
", "rawMarkdown": "In this competition dataset of lungs X-rays imgs are are taken in a fixed manner (means vertical) so here only HFlip and normalization will improve results. But you can also go with heavy augmentation it may increase your results a little bit on large models like Resnet200.", "votes": null } ], "comments": [ { "id": 1220695, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "02/28/2021 09:00:45", "content": "Adapting CutMix or MixUp to multi-label problems isn't hard, e.g. for CutMix instead of one label getting mixed, we mix all as
\n$$w_i \\times y_{kj} + (1-w_i) \\times y_{ij}$$
\nfor the jth label of image i that is having a proportion w_i of its area replaced by something cut from image k. Standard implementations in some packages may not cover it, but it's easy enough to implement.
I assume some people have experimented with it, but for me it's not been in the top of my list. The slightly tricky bit is that e.g. for CutMix, if you copy a bit of an image that does not show the anything relevant to the catheter placement, it does not actually result in a mixing of the classification of the images as far as a human is concerned. On the other hand, if you copy a bit showing a wrong placement, this changes the classification totally, not partially. It's not like photos of cats and dogs, where if you see half a dog and half a cat, it's reasonable to say 0.5 for each class. Perhaps something like SnapMix would do that better?! In any case, that's my speculation and if it ends up working well, then perhaps I'm wrong about that.
", "votes": null, "replies": [] }, { "id": 1220741, "author_name": "steamedsheep", "author_url": "", "post_date": "02/28/2021 10:07:35", "content": "Mix-based aug is usually useful but need relatively more computing resource. If we use strong mix-based aug with light transforms, the hyperparameter required to tune will reduced to few. So lazy boy like me often use these method.
", "votes": null, "replies": [] }, { "id": 1220861, "author_name": "affjljoo3581", "author_url": "", "post_date": "02/28/2021 12:25:18", "content": "I tested both Mixup and CutMix but they did not help improving performance (even they were far worse). Actually, the convergence speed was too slow. Maybe I should train with more epochs.
", "votes": null, "replies": [] }, { "id": 1221594, "author_name": "adg1822", "author_url": "", "post_date": "03/01/2021 05:41:22", "content": "In this competition dataset of lungs X-rays imgs are are taken in a fixed manner (means vertical) so here only HFlip and normalization will improve results. But you can also go with heavy augmentation it may increase your results a little bit on large models like Resnet200.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1220673": "mix-based augmentations are reported helpful in many kaggle image competitions, cutmix, mixup, fmix, etc, but I didn't see any notebook using mix-based aug, is it because the competition is multi-label and hard to adapt label?", "1220695": "Adapting CutMix or MixUp to multi-label problems isn't hard, e.g. for CutMix instead of one label getting mixed, we mix all as \n$$w_i \\times y_{kj} + (1-w_i) \\times y_{ij}$$ \nfor the jth label of image i that is having a proportion w_i of its area replaced by something cut from image k. Standard implementations in some packages may not cover it, but it's easy enough to implement.\n\nI assume some people have experimented with it, but for me it's not been in the top of my list. The slightly tricky bit is that e.g. for CutMix, if you copy a bit of an image that does not show the anything relevant to the catheter placement, it does not actually result in a mixing of the classification of the images as far as a human is concerned. On the other hand, if you copy a bit showing a wrong placement, this changes the classification totally, not partially. It's not like photos of cats and dogs, where if you see half a dog and half a cat, it's reasonable to say 0.5 for each class. Perhaps something like SnapMix would do that better?! In any case, that's my speculation and if it ends up working well, then perhaps I'm wrong about that.", "1220741": "Mix-based aug is usually useful but need relatively more computing resource. If we use strong mix-based aug with light transforms, the hyperparameter required to tune will reduced to few. So lazy boy like me often use these method.", "1220861": "I tested both Mixup and CutMix but they did not help improving performance (even they were far worse). Actually, the convergence speed was too slow. Maybe I should train with more epochs.", "1221594": "In this competition dataset of lungs X-rays imgs are are taken in a fixed manner (means vertical) so here only HFlip and normalization will improve results. But you can also go with heavy augmentation it may increase your results a little bit on large models like Resnet200." }, "source": "meta" }