This seems to be suggested by the evaluation explanations. There are very truly others images in the other dataset.
", "messages": [ { "id": 1975152, "postDate": "2022-10-06T16:15:06.760Z", "content": "My model predicts just one probability 'is this CE?' and I used images from 'other' folder with label 'Other' to enrich LAA class. Images with label 'Unknown' are not ok for this because they might have CE as well. So basically for me it was CE versus LAA + other.Other.
\nYou can see more details in my solution overview.
@pierretisseur I have tried using them with random sampling in this notebook - in the middle I have explained the usage. Here it is in short
\nFor the target labels - 'Other' and 'negative' tiles have [0,0] target, while 'LAA' tiles have [1,0] and 'CE' tiles have [0,1]. This can help avoid overfitting (to certain extent) and create a discrimination in prediction for non-critical tiles using binary cross entropy loss while training. I think using 3 outputs is not ideal as we are then forcing the network to learn something that takes away from its main task. So I use the third output as a negative output for a general case.
\nThere are some mistakes in log loss computation in the notebook (need to update), but otherwise the idea is there. Please do comment if possible what you think.
", "rawMarkdown": "@pierretisseur I have tried using them with random sampling in [this](https://www.kaggle.com/code/icemantd/feature-cluster-tile-psuedo-label-train-mayo) notebook - in the middle I have explained the usage. Here it is in short\n\nFor the target labels - 'Other' and 'negative' tiles have [0,0] target, while 'LAA' tiles have [1,0] and 'CE' tiles have [0,1]. This can help avoid overfitting (to certain extent) and create a discrimination in prediction for non-critical tiles using binary cross entropy loss while training. I think using 3 outputs is not ideal as we are then forcing the network to learn something that takes away from its main task. So I use the third output as a negative output for a general case.\n\nThere are some mistakes in log loss computation in the notebook (need to update), but otherwise the idea is there. Please do comment if possible what you think.", "votes": 2, "replies": [ { "id": 1973916, "postDate": "2022-10-06T00:15:42.027Z", "content": "Congratulation ! I would like to see your solution. The public Lb was more less a booby trap and worry me a lot since i got the worse public LB with some submission where I trusted.
", "rawMarkdown": "Congratulation ! I would like to see your solution. The public Lb was more less a booby trap and worry me a lot since i got the worse public LB with some submission where I trusted." }, { "id": 1973919, "postDate": "2022-10-06T00:20:34.583Z", "content": "Thanks @pierretisseur, can't believe what has happened!
\nCongrats on your score as well. My solution is based on above mentioned notebook, but I think I will update it properly and share full solution for getting proper feedback.
", "rawMarkdown": "Thanks @pierretisseur, can't believe what has happened!\n\nCongrats on your score as well. My solution is based on above mentioned notebook, but I think I will update it properly and share full solution for getting proper feedback. " } ] }, { "id": 1955126, "postDate": "2022-09-25T18:17:46.030Z", "content": "This seems to be suggested by the evaluation explanations. There are very truly others images in the other dataset.
", "rawMarkdown": "This seems to be suggested by the evaluation explanations. There are very truly others images in the other dataset. ", "votes": 2 } ], "comments": [ { "id": 1975152, "author_name": "Ilia los", "author_url": "", "post_date": "2022-10-06T16:15:06.760000", "content": "My model predicts just one probability 'is this CE?' and I used images from 'other' folder with label 'Other' to enrich LAA class. Images with label 'Unknown' are not ok for this because they might have CE as well. So basically for me it was CE versus LAA + other.Other.
\nYou can see more details in my solution overview.
@pierretisseur I have tried using them with random sampling in this notebook - in the middle I have explained the usage. Here it is in short
\nFor the target labels - 'Other' and 'negative' tiles have [0,0] target, while 'LAA' tiles have [1,0] and 'CE' tiles have [0,1]. This can help avoid overfitting (to certain extent) and create a discrimination in prediction for non-critical tiles using binary cross entropy loss while training. I think using 3 outputs is not ideal as we are then forcing the network to learn something that takes away from its main task. So I use the third output as a negative output for a general case.
\nThere are some mistakes in log loss computation in the notebook (need to update), but otherwise the idea is there. Please do comment if possible what you think.
", "votes": 2, "replies": [ { "id": 1973916, "author_name": "Pierre Tisseur", "author_url": "", "post_date": "2022-10-06T00:15:42.027000", "content": "Congratulation ! I would like to see your solution. The public Lb was more less a booby trap and worry me a lot since i got the worse public LB with some submission where I trusted.
", "votes": 0, "replies": [] }, { "id": 1973919, "author_name": "tdiceman", "author_url": "", "post_date": "2022-10-06T00:20:34.583000", "content": "Thanks @pierretisseur, can't believe what has happened!
\nCongrats on your score as well. My solution is based on above mentioned notebook, but I think I will update it properly and share full solution for getting proper feedback.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1975152": "My model predicts just one probability 'is this CE?' and I used images from 'other' folder with label 'Other' to enrich LAA class. Images with label 'Unknown' are not ok for this because they might have CE as well. So basically for me it was CE versus LAA + other.Other.\nYou can see more details in my [solution overview](https://www.kaggle.com/competitions/mayo-clinic-strip-ai/discussion/358089). ", "1957217": "@pierretisseur I have tried using them with random sampling in [this](https://www.kaggle.com/code/icemantd/feature-cluster-tile-psuedo-label-train-mayo) notebook - in the middle I have explained the usage. Here it is in short\n\nFor the target labels - 'Other' and 'negative' tiles have [0,0] target, while 'LAA' tiles have [1,0] and 'CE' tiles have [0,1]. This can help avoid overfitting (to certain extent) and create a discrimination in prediction for non-critical tiles using binary cross entropy loss while training. I think using 3 outputs is not ideal as we are then forcing the network to learn something that takes away from its main task. So I use the third output as a negative output for a general case.\n\nThere are some mistakes in log loss computation in the notebook (need to update), but otherwise the idea is there. Please do comment if possible what you think.", "1955126": "This seems to be suggested by the evaluation explanations. There are very truly others images in the other dataset. " } }