now both public and private score has been revealed,
\nnow both public and private score has been revealed,
\nboth ext data and pseudo label helped
", "rawMarkdown": "both ext data and pseudo label helped", "votes": 1, "replies": [ { "id": 1951260, "postDate": "2022-09-23T00:23:25.267Z", "content": "thanks. what are the score?
\nso i think i can conclude:
\n\"if used and validate properly, pseudo label always definitely helped in kaggle competition\" -- this is my experience.
\nmost of my previous gold medals are using pseudo label .
\n(for this competition, i was a bit hesitant because of additional domain shift)
for external data, it is also true but you must relabel or you must have a good aux loss.
", "rawMarkdown": "thanks. what are the score? \n\nso i think i can conclude:\n\"if used and validate properly, pseudo label always definitely helped in kaggle competition\" -- this is my experience.\nmost of my previous gold medals are using pseudo label .\n(for this competition, i was a bit hesitant because of additional domain shift)\n\nfor external data, it is also true but you must relabel or you must have a good aux loss." }, { "id": 1951262, "postDate": "2022-09-23T00:29:28.403Z", "content": "because we ensembled too much models, the improvement of pseudo label on ext data is smaller than 0.01. I don't know the exact score
", "rawMarkdown": "because we ensembled too much models, the improvement of pseudo label on ext data is smaller than 0.01. I don't know the exact score", "votes": 1 }, { "id": 1952009, "postDate": "2022-09-23T11:38:38.027Z", "content": "Yes, we use both HPA and HUBMAP external data to improve, when our single-mode lb_hubmap reaches around 0.60.
\nUsing HPA's external data boost LB_hubmap around 0.005+,and data from hpa data;
\nHuBMAP's external data boost LB_hubmap around 0.005-, and data from hubmap data.
external data helped a lot, as well as pseudo
", "rawMarkdown": "external data helped a lot, as well as pseudo ", "votes": 2, "replies": [ { "id": 1951721, "postDate": "2022-09-23T07:44:59.010Z", "content": "I have collected a lot of medical images but I don't know how to use.
", "rawMarkdown": "I have collected a lot of medical images but I don't know how to use." }, { "id": 1952409, "postDate": "2022-09-23T16:42:34.673Z", "content": "We collected around 250k images from the human protein atlas and GTEX portals, pseudo-labeled it and during training randomly sampled a subset.
", "rawMarkdown": "We collected around 250k images from the human protein atlas and GTEX portals, pseudo-labeled it and during training randomly sampled a subset. \n\n\n", "votes": 3 }, { "id": 1952462, "postDate": "2022-09-23T17:21:39.567Z", "content": "Thanks for sharing!
", "rawMarkdown": "Thanks for sharing! " } ] }, { "id": 1951462, "postDate": "2022-09-23T04:41:09.120Z", "content": "I used a lot of external data and I really don't know it helped or not. I used pseudo labels of 25 prostate and 25 spleen slices from GTEx whole slide images, HuBMAP colonic crypt dataset and pseudo labels of single test image. I got the worst shake down in top 50.
", "rawMarkdown": "I used a lot of external data and I really don't know it helped or not. I used pseudo labels of 25 prostate and 25 spleen slices from GTEx whole slide images, HuBMAP colonic crypt dataset and pseudo labels of single test image. I got the worst shake down in top 50." }, { "id": 1951444, "postDate": "2022-09-23T04:20:35.087Z", "content": "Focus on backbone and decoder:
\nBackbone: Swin-base< nextvit-large <convnext-large<pvtv2-b4<coat-lite-medium
\nwithout externel data and relabel, public lb=0.81, prostate=0.19, private lb=0.79, prostate=0.17
\nFor prostate, try SegNeXt which replace self-attention with large conv kernel, however it is easy to overfit.
Decoder: dafomer is best, replace stage-3/4 with unext, public lb=0.81.
\nHave no time to try relabel and externel data, may be it is usefult
", "rawMarkdown": "Focus on backbone and decoder:\n\nBackbone: Swin-base< nextvit-largeprivate: baseline + lung-relabel > baseline + lung-relabel + prostate-relabel (both 0.80)
\nI expect relabeling prostate is more useful, but is not.
", "rawMarkdown": "public: baseline + lung-relabel < baseline + lung-relabel + prostate-relabel (both 0.81)\n\nprivate: baseline + lung-relabel > baseline + lung-relabel + prostate-relabel (both 0.80)\n\nI expect relabeling prostate is more useful, but is not.", "replies": [ { "id": 1951404, "postDate": "2022-09-23T03:43:40.360Z", "content": "Yes, I got similar result.
", "rawMarkdown": "Yes, I got similar result." }, { "id": 1951483, "postDate": "2022-09-23T05:09:09.877Z", "content": "public of lung: baseline(0.11), lung-relabel(0.11)
\nprivate of lung: baseline(0.18), lung-relabel(0.19)
", "rawMarkdown": "public of lung: baseline(0.11), lung-relabel(0.11) \n\nprivate of lung: baseline(0.18), lung-relabel(0.19)" } ] }, { "id": 1951332, "postDate": "2022-09-23T02:37:43.527Z", "content": "We have tried some expriments like mae trained on some medical images we collected, but it performs poorly. Maybe I don't have a good training strategy?
", "rawMarkdown": "We have tried some expriments like mae trained on some medical images we collected, but it performs poorly. Maybe I don't have a good training strategy?" }, { "id": 1951310, "postDate": "2022-09-23T02:17:32.607Z", "content": "https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/354590
", "rawMarkdown": "https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/354590" }, { "id": 1951271, "postDate": "2022-09-23T00:59:12.237Z", "content": "\"did anyone succeed with pseudo labeling? \"
\nYes, we used the pseudo label for lung which help us reach 0.81 from 0.80 on the public LB when our backbone is mit-b2 and our image size is 768*768!
Can you share your lung score on public LB and private LB?
", "rawMarkdown": "Can you share your lung score on public LB and private LB?" }, { "id": 1951273, "postDate": "2022-09-23T01:00:55.753Z", "content": "thanks.
\nhow about private score?
", "rawMarkdown": "thanks.\n\nhow about private score?" } ] }, { "id": 1951291, "postDate": "2022-09-23T01:22:17.477Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 1951307, "postDate": "2022-09-23T02:13:11.947Z", "content": "You should just use the pseudo label for lung which really works well.
", "rawMarkdown": "You should just use the pseudo label for lung which really works well.", "votes": 1 } ] } ], "comments": [ { "id": 1951257, "author_name": "Carno Zhao", "author_url": "", "post_date": "2022-09-23T00:21:28.920000", "content": "both ext data and pseudo label helped
", "votes": 1, "replies": [ { "id": 1951260, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-23T00:23:25.267000", "content": "thanks. what are the score?
\nso i think i can conclude:
\n\"if used and validate properly, pseudo label always definitely helped in kaggle competition\" -- this is my experience.
\nmost of my previous gold medals are using pseudo label .
\n(for this competition, i was a bit hesitant because of additional domain shift)
for external data, it is also true but you must relabel or you must have a good aux loss.
", "votes": 0, "replies": [] }, { "id": 1951262, "author_name": "Carno Zhao", "author_url": "", "post_date": "2022-09-23T00:29:28.403000", "content": "because we ensembled too much models, the improvement of pseudo label on ext data is smaller than 0.01. I don't know the exact score
", "votes": 1, "replies": [] }, { "id": 1952009, "author_name": "Roc", "author_url": "", "post_date": "2022-09-23T11:38:38.027000", "content": "Yes, we use both HPA and HUBMAP external data to improve, when our single-mode lb_hubmap reaches around 0.60.
\nUsing HPA's external data boost LB_hubmap around 0.005+,and data from hpa data;
\nHuBMAP's external data boost LB_hubmap around 0.005-, and data from hubmap data.
external data helped a lot, as well as pseudo
", "votes": 2, "replies": [ { "id": 1951721, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T07:44:59.010000", "content": "I have collected a lot of medical images but I don't know how to use.
", "votes": 0, "replies": [] }, { "id": 1952409, "author_name": "Igor Krashenyi", "author_url": "", "post_date": "2022-09-23T16:42:34.673000", "content": "We collected around 250k images from the human protein atlas and GTEX portals, pseudo-labeled it and during training randomly sampled a subset.
", "votes": 3, "replies": [] }, { "id": 1952462, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T17:21:39.567000", "content": "Thanks for sharing!
", "votes": 0, "replies": [] } ] }, { "id": 1951462, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2022-09-23T04:41:09.120000", "content": "I used a lot of external data and I really don't know it helped or not. I used pseudo labels of 25 prostate and 25 spleen slices from GTEx whole slide images, HuBMAP colonic crypt dataset and pseudo labels of single test image. I got the worst shake down in top 50.
", "votes": 0, "replies": [] }, { "id": 1951444, "author_name": "跟着丞相割麦子", "author_url": "", "post_date": "2022-09-23T04:20:35.087000", "content": "Focus on backbone and decoder:
\nBackbone: Swin-base< nextvit-large <convnext-large<pvtv2-b4<coat-lite-medium
\nwithout externel data and relabel, public lb=0.81, prostate=0.19, private lb=0.79, prostate=0.17
\nFor prostate, try SegNeXt which replace self-attention with large conv kernel, however it is easy to overfit.
Decoder: dafomer is best, replace stage-3/4 with unext, public lb=0.81.
\nHave no time to try relabel and externel data, may be it is usefult
", "votes": 0, "replies": [] }, { "id": 1951393, "author_name": "gray98", "author_url": "", "post_date": "2022-09-23T03:25:16.420000", "content": "public: baseline + lung-relabel < baseline + lung-relabel + prostate-relabel (both 0.81)
\nprivate: baseline + lung-relabel > baseline + lung-relabel + prostate-relabel (both 0.80)
\nI expect relabeling prostate is more useful, but is not.
", "votes": 0, "replies": [ { "id": 1951404, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T03:43:40.360000", "content": "Yes, I got similar result.
", "votes": 0, "replies": [] }, { "id": 1951483, "author_name": "gray98", "author_url": "", "post_date": "2022-09-23T05:09:09.877000", "content": "public of lung: baseline(0.11), lung-relabel(0.11)
\nprivate of lung: baseline(0.18), lung-relabel(0.19)
", "votes": 0, "replies": [] } ] }, { "id": 1951332, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T02:37:43.527000", "content": "We have tried some expriments like mae trained on some medical images we collected, but it performs poorly. Maybe I don't have a good training strategy?
", "votes": 0, "replies": [] }, { "id": 1951310, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T02:17:32.607000", "content": "https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/354590
", "votes": 0, "replies": [] }, { "id": 1951271, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T00:59:12.237000", "content": "\"did anyone succeed with pseudo labeling? \"
\nYes, we used the pseudo label for lung which help us reach 0.81 from 0.80 on the public LB when our backbone is mit-b2 and our image size is 768*768!
Can you share your lung score on public LB and private LB?
", "votes": 0, "replies": [] }, { "id": 1951273, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-23T01:00:55.753000", "content": "thanks.
\nhow about private score?
", "votes": 0, "replies": [] } ] }, { "id": 1951291, "author_name": "", "author_url": "", "post_date": "2022-09-23T01:22:17.477000", "content": "", "votes": 0, "replies": [ { "id": 1951307, "author_name": "Chinartist", "author_url": "", "post_date": "2022-09-23T02:13:11.947000", "content": "You should just use the pseudo label for lung which really works well.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1951255": "now both public and private score has been revealed,\n\n0. did anyone succeed with external data? i.e. both public and private score are improved\n1. did anyone succeed with pseudo labeling? i.e. both public and private score are improved\n2. did anyone succeed with transformer pre-training? e.g. masked image modelling\n3. did any of the semi-supervised methods worked? e.g. mean teacher\n\n---\n\n4. Swin trasnformer (v1,v2,v3) performs poorly (single model with ensemble is below 0.78). did anyone gets good results?\n", "1951257": "both ext data and pseudo label helped", "1951689": "external data helped a lot, as well as pseudo ", "1951462": "I used a lot of external data and I really don't know it helped or not. I used pseudo labels of 25 prostate and 25 spleen slices from GTEx whole slide images, HuBMAP colonic crypt dataset and pseudo labels of single test image. I got the worst shake down in top 50.", "1951444": "Focus on backbone and decoder:\n\nBackbone: Swin-base< nextvit-large