Hi all,
\n\nCongratulations to the winners!
\n\nThere were two items that worked that I haven't seen mentioned thus far:
\n\nMixed slide augmentations: I got my highest CV scores when using these (and they also generated some of the highest private LB scores). When an image was selected for training I randomly selected a second image from the same data provider, same isup grade, tiled both images, then randomly selected 50% of the tiles from each (and reordered them as you would if you were selecting tiles from a single image). This lead to a +0.005 in CV score. I tried a version whereby , as well as the same data provider + isup, the mix image was of the same tissue colour, but this didn't add anything additional.
Sampling according to the Karolinska distribution: before each epoch the Radboud images were sampled according to the Karolinska distribution of isup grades (i.e. sample proportion of grades 0 - 5 as Karolinska). The idea was that the Karolinska distribution was more representative of the test set.
Cheers
", "messages": [ { "id": 943220, "postDate": "2020-07-24T08:00:22.193Z", "content": "Hi all,
\n\nCongratulations to the winners!
\n\nThere were two items that worked that I haven't seen mentioned thus far:
\n\nMixed slide augmentations: I got my highest CV scores when using these (and they also generated some of the highest private LB scores). When an image was selected for training I randomly selected a second image from the same data provider, same isup grade, tiled both images, then randomly selected 50% of the tiles from each (and reordered them as you would if you were selecting tiles from a single image). This lead to a +0.005 in CV score. I tried a version whereby , as well as the same data provider + isup, the mix image was of the same tissue colour, but this didn't add anything additional.
Sampling according to the Karolinska distribution: before each epoch the Radboud images were sampled according to the Karolinska distribution of isup grades (i.e. sample proportion of grades 0 - 5 as Karolinska). The idea was that the Karolinska distribution was more representative of the test set.
Cheers
", "rawMarkdown": "Hi all,\n\nCongratulations to the winners! \n\nThere were two items that worked that I haven't seen mentioned thus far:\n\n1. Mixed slide augmentations: I got my highest CV scores when using these (and they also generated some of the highest private LB scores). When an image was selected for training I randomly selected a second image from the same data provider, same isup grade, tiled both images, then randomly selected 50% of the tiles from each (and reordered them as you would if you were selecting tiles from a single image). This lead to a +0.005 in CV score. I tried a version whereby , as well as the same data provider + isup, the mix image was of the same tissue colour, but this didn't add anything additional.\n\n2. Sampling according to the Karolinska distribution: before each epoch the Radboud images were sampled according to the Karolinska distribution of isup grades (i.e. sample proportion of grades 0 - 5 as Karolinska). The idea was that the Karolinska distribution was more representative of the test set. \n\nCheers", "votes": 11 }, { "id": 943854, "postDate": "2020-07-24T16:16:58.547Z", "content": "Hi thanks for sharing.\nI tried a similar cutmix augmentation but more \"permissive\" : my second image was randomly selected among all the images i.e. not always the same data provider/isup grade. I then changed the label based on the proportion of image 1 and 2 (I randomly select the number of tiles). So labels could be 3.x, 0.x ... I've observed a small decrease in CV so I did not push forward the idea
", "rawMarkdown": "Hi thanks for sharing.\nI tried a similar cutmix augmentation but more \"permissive\" : my second image was randomly selected among all the images i.e. not always the same data provider/isup grade. I then changed the label based on the proportion of image 1 and 2 (I randomly select the number of tiles). So labels could be 3.x, 0.x ... I've observed a small decrease in CV so I did not push forward the idea", "votes": 1, "replies": [ { "id": 943868, "postDate": "2020-07-24T16:30:52.593Z", "content": "So you mixed the images and then tiled the mixed image? Or tiled them individually and then mixed the tiles like I did?
", "rawMarkdown": "So you mixed the images and then tiled the mixed image? Or tiled them individually and then mixed the tiles like I did?" }, { "id": 943913, "postDate": "2020-07-24T17:04:53.610Z", "content": "Like you did
", "rawMarkdown": "Like you did", "votes": 1 } ] }, { "id": 947866, "postDate": "2020-07-27T14:25:51.147Z", "content": "Thanks for sharing. How did you get the idea to do it?
", "rawMarkdown": "Thanks for sharing. How did you get the idea to do it?", "replies": [ { "id": 947877, "postDate": "2020-07-27T14:31:39.070Z", "content": "Just thinking about how best to generalise given the limited data!
", "rawMarkdown": "Just thinking about how best to generalise given the limited data!", "votes": 1 }, { "id": 947897, "postDate": "2020-07-27T14:43:24.827Z", "content": "Is it like applying MixUp to individual tiles?
", "rawMarkdown": "Is it like applying MixUp to individual tiles?" }, { "id": 947960, "postDate": "2020-07-27T15:17:06.447Z", "content": "It would be a similar to mixup, but we're keeping the tiles intact (i.e. the individual tiles are not a mix from two images), but then mixing the tile selections from two images
", "rawMarkdown": "It would be a similar to mixup, but we're keeping the tiles intact (i.e. the individual tiles are not a mix from two images), but then mixing the tile selections from two images" } ] }, { "id": 943327, "postDate": "2020-07-24T09:48:20.163Z", "content": "Thanks for sharing @fergusoci
", "rawMarkdown": "Thanks for sharing @fergusoci ", "votes": 1 } ], "comments": [ { "id": 943854, "author_name": "Alex", "author_url": "", "post_date": "2020-07-24T16:16:58.547000", "content": "Hi thanks for sharing.\nI tried a similar cutmix augmentation but more \"permissive\" : my second image was randomly selected among all the images i.e. not always the same data provider/isup grade. I then changed the label based on the proportion of image 1 and 2 (I randomly select the number of tiles). So labels could be 3.x, 0.x ... I've observed a small decrease in CV so I did not push forward the idea
", "votes": 1, "replies": [ { "id": 943868, "author_name": "fergusoci", "author_url": "", "post_date": "2020-07-24T16:30:52.593000", "content": "So you mixed the images and then tiled the mixed image? Or tiled them individually and then mixed the tiles like I did?
", "votes": 0, "replies": [] }, { "id": 943913, "author_name": "Alex", "author_url": "", "post_date": "2020-07-24T17:04:53.610000", "content": "Like you did
", "votes": 1, "replies": [] } ] }, { "id": 947866, "author_name": "Son of Anton v3.0", "author_url": "", "post_date": "2020-07-27T14:25:51.147000", "content": "Thanks for sharing. How did you get the idea to do it?
", "votes": 0, "replies": [ { "id": 947877, "author_name": "fergusoci", "author_url": "", "post_date": "2020-07-27T14:31:39.070000", "content": "Just thinking about how best to generalise given the limited data!
", "votes": 1, "replies": [] }, { "id": 947897, "author_name": "Son of Anton v3.0", "author_url": "", "post_date": "2020-07-27T14:43:24.827000", "content": "Is it like applying MixUp to individual tiles?
", "votes": 0, "replies": [] }, { "id": 947960, "author_name": "fergusoci", "author_url": "", "post_date": "2020-07-27T15:17:06.447000", "content": "It would be a similar to mixup, but we're keeping the tiles intact (i.e. the individual tiles are not a mix from two images), but then mixing the tile selections from two images
", "votes": 0, "replies": [] } ] }, { "id": 943327, "author_name": "Gryffindor", "author_url": "", "post_date": "2020-07-24T09:48:20.163000", "content": "Thanks for sharing @fergusoci
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "943220": "Hi all,\n\nCongratulations to the winners! \n\nThere were two items that worked that I haven't seen mentioned thus far:\n\n1. Mixed slide augmentations: I got my highest CV scores when using these (and they also generated some of the highest private LB scores). When an image was selected for training I randomly selected a second image from the same data provider, same isup grade, tiled both images, then randomly selected 50% of the tiles from each (and reordered them as you would if you were selecting tiles from a single image). This lead to a +0.005 in CV score. I tried a version whereby , as well as the same data provider + isup, the mix image was of the same tissue colour, but this didn't add anything additional.\n\n2. Sampling according to the Karolinska distribution: before each epoch the Radboud images were sampled according to the Karolinska distribution of isup grades (i.e. sample proportion of grades 0 - 5 as Karolinska). The idea was that the Karolinska distribution was more representative of the test set. \n\nCheers", "943854": "Hi thanks for sharing.\nI tried a similar cutmix augmentation but more \"permissive\" : my second image was randomly selected among all the images i.e. not always the same data provider/isup grade. I then changed the label based on the proportion of image 1 and 2 (I randomly select the number of tiles). So labels could be 3.x, 0.x ... I've observed a small decrease in CV so I did not push forward the idea", "947866": "Thanks for sharing. How did you get the idea to do it?", "943327": "Thanks for sharing @fergusoci " } }