I took a quick look in the training data and the ETT targets are exclusive, so one could treat them as a multiclass classifiction by adding \"no ETT\" as a 4th ETT class. Did anyone try this?
\nOn the other hand, NGT can be, for example, normal and abnormal at the same time and CVC can be, for example, borderline and normal at the same time. Can anyone explain why it is like this?
", "messages": [ { "id": 1137319, "postDate": "2021-01-03T20:44:04.560Z", "content": "The reason is, that only single ETT is placed in a patient (so either there is ETT or there is not).
\nAs for CVC and NGT - multiple of the tubes can be inserted. It is not uncommon to see 3 CVC's inserted!
", "rawMarkdown": "The reason is, that only single ETT is placed in a patient (so either there is ETT or there is not).\n\nAs for CVC and NGT - multiple of the tubes can be inserted. It is not uncommon to see 3 CVC's inserted!", "votes": 6 }, { "id": 1137095, "postDate": "2021-01-03T17:14:03.787Z", "content": "I added 'no ETT' class and used softmax for 4 ETT classes (sigmoid for other 8 classes).
\nAlthough my CV was improved, public LB score wasn't.
Your question might be answered in https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203372 by @jarrelscy.
", "rawMarkdown": "I added 'no ETT' class and used softmax for 4 ETT classes (sigmoid for other 8 classes).\nAlthough my CV was improved, public LB score wasn't.\n\nYour question might be answered in https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203372 by @jarrelscy.", "votes": 3, "replies": [ { "id": 1137162, "postDate": "2021-01-03T18:16:49.237Z", "content": "I thought to do exactly that.
\nAlso, thanks for the link!
", "rawMarkdown": "I thought to do exactly that. \n\nAlso, thanks for the link!" }, { "id": 1139219, "postDate": "2021-01-05T08:53:51.203Z", "content": "\"Although my CV was improved, public LB score wasn't.\"
\nthere is a possibility that private LB improve, if we assume cv = (public+private)/2
\nbeing the easiest class, but I think most people already have the perfect ROC curve for ETT case even when using sigmoid. so the improvement may not be visible?
", "rawMarkdown": "\"Although my CV was improved, public LB score wasn't.\"\n\nthere is a possibility that private LB improve, if we assume cv = (public+private)/2\n\nbeing the easiest class, but I think most people already have the perfect ROC curve for ETT case even when using sigmoid. so the improvement may not be visible?", "votes": 2 }, { "id": 1149706, "postDate": "2021-01-12T04:40:25.553Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 1149765, "postDate": "2021-01-12T05:59:27.160Z", "content": "Thank you for your question!
\nSimply, I used bce for all classes.
I took a quick look in the training data and the ETT targets are exclusive, so one could treat them as a multiclass classifiction by adding \"no ETT\" as a 4th ETT class. Did anyone try this?
\nOn the other hand, NGT can be, for example, normal and abnormal at the same time and CVC can be, for example, borderline and normal at the same time. Can anyone explain why it is like this?
", "rawMarkdown": "I took a quick look in the training data and the ETT targets are exclusive, so one could treat them as a multiclass classifiction by adding \"no ETT\" as a 4th ETT class. Did anyone try this? \n\nOn the other hand, NGT can be, for example, normal and abnormal at the same time and CVC can be, for example, borderline and normal at the same time. Can anyone explain why it is like this?", "votes": 4 } ], "comments": [ { "id": 1137319, "author_name": "raddar", "author_url": "", "post_date": "2021-01-03T20:44:04.560000", "content": "The reason is, that only single ETT is placed in a patient (so either there is ETT or there is not).
\nAs for CVC and NGT - multiple of the tubes can be inserted. It is not uncommon to see 3 CVC's inserted!
", "votes": 6, "replies": [] }, { "id": 1137095, "author_name": "YYama", "author_url": "", "post_date": "2021-01-03T17:14:03.787000", "content": "I added 'no ETT' class and used softmax for 4 ETT classes (sigmoid for other 8 classes).
\nAlthough my CV was improved, public LB score wasn't.
Your question might be answered in https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203372 by @jarrelscy.
", "votes": 3, "replies": [ { "id": 1137162, "author_name": "Marcel", "author_url": "", "post_date": "2021-01-03T18:16:49.237000", "content": "I thought to do exactly that.
\nAlso, thanks for the link!
", "votes": 0, "replies": [] }, { "id": 1139219, "author_name": "hengck23", "author_url": "", "post_date": "2021-01-05T08:53:51.203000", "content": "\"Although my CV was improved, public LB score wasn't.\"
\nthere is a possibility that private LB improve, if we assume cv = (public+private)/2
\nbeing the easiest class, but I think most people already have the perfect ROC curve for ETT case even when using sigmoid. so the improvement may not be visible?
", "votes": 2, "replies": [] }, { "id": 1149706, "author_name": "", "author_url": "", "post_date": "2021-01-12T04:40:25.553000", "content": "", "votes": 1, "replies": [] }, { "id": 1149765, "author_name": "YYama", "author_url": "", "post_date": "2021-01-12T05:59:27.160000", "content": "Thank you for your question!
\nSimply, I used bce for all classes.