{ "id": 464186, "title": "Highest score without predicting Other", "url": "/competitions/UBC-OCEAN/discussion/464186", "author_name": "Gunes Evitan", "post_date": "2023-12-29T05:18:34.665000", "votes": 16, "comment_count": 18, "views": 0, "content": "

I was stuck at high 0.47 for a while and decided to explore outlier prediction part of this competition. I started predicting Other's and my LB score jumped from 0.47 to 0.54. I was wondering if my model still has any room for improvement for regular classes. What is your highest score without predicting Other?

", "messages": [ { "id": 2578200, "postDate": "2023-12-29T05:18:34.667Z", "content": "

I was stuck at high 0.47 for a while and decided to explore outlier prediction part of this competition. I started predicting Other's and my LB score jumped from 0.47 to 0.54. I was wondering if my model still has any room for improvement for regular classes. What is your highest score without predicting Other?

", "rawMarkdown": "I was stuck at high 0.47 for a while and decided to explore outlier prediction part of this competition. I started predicting Other's and my LB score jumped from 0.47 to 0.54. I was wondering if my model still has any room for improvement for regular classes. What is your highest score without predicting Other?", "votes": 16 }, { "id": 2578246, "postDate": "2023-12-29T05:57:35.967Z", "content": "

My current LB score was a 5-fold v2s trained only on Kaggle data without predicting Other. The masks provided by the host were pretty useless 😅. External datasets are definitely needed in order to train models which are capable of predicting Other

", "rawMarkdown": "My current LB score was a 5-fold v2s trained only on Kaggle data without predicting Other. The masks provided by the host were pretty useless 😅. External datasets are definitely needed in order to train models which are capable of predicting Other", "votes": 6, "replies": [ { "id": 2578254, "postDate": "2023-12-29T06:01:14.723Z", "content": "

Wow your score is too good. I guess you are not even using any localization at all?

", "rawMarkdown": "Wow your score is too good. I guess you are not even using any localization at all?", "replies": [ { "id": 2578322, "postDate": "2023-12-29T06:36:21.197Z", "content": "

yup, I applied basic image processing to filter non-tissue regions on each slide. Also, there was a little trick to speed up tiling. My models used 16 1024x1024 tiles per slide.

", "rawMarkdown": "yup, I applied basic image processing to filter non-tissue regions on each slide. Also, there was a little trick to speed up tiling. My models used 16 1024x1024 tiles per slide.", "votes": 6 } ] }, { "id": 2582936, "postDate": "2024-01-02T00:27:38.550Z", "content": "

Yeah, the quality of provided mask is really bad…

", "rawMarkdown": "Yeah, the quality of provided mask is really bad...", "votes": 1 } ] }, { "id": 2586068, "postDate": "2024-01-04T00:33:08.840Z", "content": "

Without outlier detection, our team's highest score on the public is 0.6, and the highest score on the private is 0.58 (no additional dataset is needed)

", "rawMarkdown": "Without outlier detection, our team's highest score on the public is 0.6, and the highest score on the private is 0.58 (no additional dataset is needed)", "votes": 4, "replies": [ { "id": 2586071, "postDate": "2024-01-04T00:37:42.973Z", "content": "

What is your only TMA private score? My only private TMA is 0.33. Meaning, my private WSI score is low.

", "rawMarkdown": "What is your only TMA private score? My only private TMA is 0.33. Meaning, my private WSI score is low.", "replies": [ { "id": 2586073, "postDate": "2024-01-04T00:43:10.967Z", "content": "

I haven't tested this yet. I will test it when I write the solution. My experience is that the score between TMA and WSI are close.

", "rawMarkdown": "I haven't tested this yet. I will test it when I write the solution. My experience is that the score between TMA and WSI are close." } ] }, { "id": 2586265, "postDate": "2024-01-04T05:06:59.893Z", "content": "

Very impressive score without external data.

", "rawMarkdown": "Very impressive score without external data." } ] }, { "id": 2584794, "postDate": "2024-01-03T07:37:53.037Z", "content": "

My score (0.54) is also the score without using 'Other'. I’m excited to hear other participants’ outlier detection solutions in a few hours!

", "rawMarkdown": "My score (0.54) is also the score without using 'Other'. I’m excited to hear other participants’ outlier detection solutions in a few hours!", "votes": 1, "replies": [ { "id": 2584802, "postDate": "2024-01-03T07:40:15.120Z", "content": "

You will be very disappointed when you hear how I do it.

", "rawMarkdown": "You will be very disappointed when you hear how I do it.", "votes": 1, "replies": [ { "id": 2584815, "postDate": "2024-01-03T07:47:22.390Z", "content": "

Sometimes simple is the best😭

", "rawMarkdown": "Sometimes simple is the best😭", "votes": 1 } ] } ] }, { "id": 2585597, "postDate": "2024-01-03T16:52:06.157Z", "content": "

My score 0.52 without using 'Other'. I process Other class, by classify Tumor or Non tumor tile, but not increase score :((

", "rawMarkdown": "My score 0.52 without using 'Other'. I process Other class, by classify Tumor or Non tumor tile, but not increase score :((", "votes": 2 }, { "id": 2578394, "postDate": "2023-12-29T08:24:21.260Z", "content": "

My best score (0.5) was achieved with a model that completely ignores the Other class. Theoretically we should be able to reach 5/6=~83% so I tried a lot of variants but none of them reached the 50%.
\nI think the only way to further improve the 50% is to include outlier detection.

", "rawMarkdown": "My best score (0.5) was achieved with a model that completely ignores the Other class. Theoretically we should be able to reach 5/6=~83% so I tried a lot of variants but none of them reached the 50%.\nI think the only way to further improve the 50% is to include outlier detection.", "votes": 2 }, { "id": 2578223, "postDate": "2023-12-29T05:41:31.677Z", "content": "

Any tips on how you are predicting the Other label? I tried training my model with different histological images from different datasets for the \"Other\" label but that doesnt seem to be working well. As well as tried adding a filter that classifies an image as Other if the other class probabilities are close enough

", "rawMarkdown": "Any tips on how you are predicting the Other label? I tried training my model with different histological images from different datasets for the \"Other\" label but that doesnt seem to be working well. As well as tried adding a filter that classifies an image as Other if the other class probabilities are close enough", "votes": 2, "replies": [ { "id": 2578251, "postDate": "2023-12-29T05:59:34.133Z", "content": "

Sorry, I can't disclose that information since we are in the final week.

", "rawMarkdown": "Sorry, I can't disclose that information since we are in the final week.", "votes": 2 } ] }, { "id": 2578355, "postDate": "2023-12-29T07:16:27.593Z", "content": "

Your accuracy for predicting 'Others' is great! Besides, though details cannot be revealed before the end of the challenge, I believe the highest score without 'Other' can be ~0.55 as commented above.

", "rawMarkdown": "Your accuracy for predicting 'Others' is great! Besides, though details cannot be revealed before the end of the challenge, I believe the highest score without 'Other' can be ~0.55 as commented above.", "replies": [ { "id": 2578366, "postDate": "2023-12-29T07:26:31.353Z", "content": "

Thanks… I hope the private test set will have more Others so I can jump to the gold zone :D

", "rawMarkdown": "Thanks... I hope the private test set will have more Others so I can jump to the gold zone :D", "replies": [ { "id": 2586266, "postDate": "2024-01-04T05:07:44.420Z", "content": "

and that just happened

", "rawMarkdown": "and that just happened" } ] } ] } ], "comments": [ { "id": 2578246, "author_name": "NguyenThanhNhan", "author_url": "", "post_date": "2023-12-29T05:57:35.967000", "content": "

My current LB score was a 5-fold v2s trained only on Kaggle data without predicting Other. The masks provided by the host were pretty useless 😅. External datasets are definitely needed in order to train models which are capable of predicting Other

", "votes": 6, "replies": [ { "id": 2578254, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-12-29T06:01:14.723000", "content": "

Wow your score is too good. I guess you are not even using any localization at all?

", "votes": 0, "replies": [ { "id": 2578322, "author_name": "NguyenThanhNhan", "author_url": "", "post_date": "2023-12-29T06:36:21.197000", "content": "

yup, I applied basic image processing to filter non-tissue regions on each slide. Also, there was a little trick to speed up tiling. My models used 16 1024x1024 tiles per slide.

", "votes": 6, "replies": [] } ] }, { "id": 2582936, "author_name": "Jun Huang", "author_url": "", "post_date": "2024-01-02T00:27:38.550000", "content": "

Yeah, the quality of provided mask is really bad…

", "votes": 1, "replies": [] } ] }, { "id": 2586068, "author_name": "m1dsolo", "author_url": "", "post_date": "2024-01-04T00:33:08.840000", "content": "

Without outlier detection, our team's highest score on the public is 0.6, and the highest score on the private is 0.58 (no additional dataset is needed)

", "votes": 4, "replies": [ { "id": 2586071, "author_name": "Quan Vu", "author_url": "", "post_date": "2024-01-04T00:37:42.973000", "content": "

What is your only TMA private score? My only private TMA is 0.33. Meaning, my private WSI score is low.

", "votes": 0, "replies": [ { "id": 2586073, "author_name": "m1dsolo", "author_url": "", "post_date": "2024-01-04T00:43:10.967000", "content": "

I haven't tested this yet. I will test it when I write the solution. My experience is that the score between TMA and WSI are close.

", "votes": 0, "replies": [] } ] }, { "id": 2586265, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2024-01-04T05:06:59.893000", "content": "

Very impressive score without external data.

", "votes": 0, "replies": [] } ] }, { "id": 2584794, "author_name": "devchopin", "author_url": "", "post_date": "2024-01-03T07:37:53.037000", "content": "

My score (0.54) is also the score without using 'Other'. I’m excited to hear other participants’ outlier detection solutions in a few hours!

", "votes": 1, "replies": [ { "id": 2584802, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2024-01-03T07:40:15.120000", "content": "

You will be very disappointed when you hear how I do it.

", "votes": 1, "replies": [ { "id": 2584815, "author_name": "devchopin", "author_url": "", "post_date": "2024-01-03T07:47:22.390000", "content": "

Sometimes simple is the best😭

", "votes": 1, "replies": [] } ] } ] }, { "id": 2585597, "author_name": "Quan Vu", "author_url": "", "post_date": "2024-01-03T16:52:06.157000", "content": "

My score 0.52 without using 'Other'. I process Other class, by classify Tumor or Non tumor tile, but not increase score :((

", "votes": 2, "replies": [] }, { "id": 2578394, "author_name": "Manuel K", "author_url": "", "post_date": "2023-12-29T08:24:21.260000", "content": "

My best score (0.5) was achieved with a model that completely ignores the Other class. Theoretically we should be able to reach 5/6=~83% so I tried a lot of variants but none of them reached the 50%.
\nI think the only way to further improve the 50% is to include outlier detection.

", "votes": 2, "replies": [] }, { "id": 2578223, "author_name": "pjmathematician", "author_url": "", "post_date": "2023-12-29T05:41:31.677000", "content": "

Any tips on how you are predicting the Other label? I tried training my model with different histological images from different datasets for the \"Other\" label but that doesnt seem to be working well. As well as tried adding a filter that classifies an image as Other if the other class probabilities are close enough

", "votes": 2, "replies": [ { "id": 2578251, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-12-29T05:59:34.133000", "content": "

Sorry, I can't disclose that information since we are in the final week.

", "votes": 2, "replies": [] } ] }, { "id": 2578355, "author_name": "Zijie Fang", "author_url": "", "post_date": "2023-12-29T07:16:27.593000", "content": "

Your accuracy for predicting 'Others' is great! Besides, though details cannot be revealed before the end of the challenge, I believe the highest score without 'Other' can be ~0.55 as commented above.

", "votes": 0, "replies": [ { "id": 2578366, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-12-29T07:26:31.353000", "content": "

Thanks… I hope the private test set will have more Others so I can jump to the gold zone :D

", "votes": 0, "replies": [ { "id": 2586266, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2024-01-04T05:07:44.420000", "content": "

and that just happened

", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2578200": "I was stuck at high 0.47 for a while and decided to explore outlier prediction part of this competition. I started predicting Other's and my LB score jumped from 0.47 to 0.54. I was wondering if my model still has any room for improvement for regular classes. What is your highest score without predicting Other?", "2578246": "My current LB score was a 5-fold v2s trained only on Kaggle data without predicting Other. The masks provided by the host were pretty useless 😅. External datasets are definitely needed in order to train models which are capable of predicting Other", "2586068": "Without outlier detection, our team's highest score on the public is 0.6, and the highest score on the private is 0.58 (no additional dataset is needed)", "2584794": "My score (0.54) is also the score without using 'Other'. I’m excited to hear other participants’ outlier detection solutions in a few hours!", "2585597": "My score 0.52 without using 'Other'. I process Other class, by classify Tumor or Non tumor tile, but not increase score :((", "2578394": "My best score (0.5) was achieved with a model that completely ignores the Other class. Theoretically we should be able to reach 5/6=~83% so I tried a lot of variants but none of them reached the 50%.\nI think the only way to further improve the 50% is to include outlier detection.", "2578223": "Any tips on how you are predicting the Other label? I tried training my model with different histological images from different datasets for the \"Other\" label but that doesnt seem to be working well. As well as tried adding a filter that classifies an image as Other if the other class probabilities are close enough", "2578355": "Your accuracy for predicting 'Others' is great! Besides, though details cannot be revealed before the end of the challenge, I believe the highest score without 'Other' can be ~0.55 as commented above." } }