\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3317928%2F8b6f053109e1cac6440882f3f0fd3ce0%2Fprobs.png?generation=1691483480185093&alt=media\")
Hi all!
\nAs we approach the competition conclusion, Everyone are attempting to move up the leaderboard as the game draws to a close. To determine the ideal threshold value, I am now doing experiments. Lower thresholds are recommended by publicly accessible code and discussion. So I try 0.45 or 0.40, but neither of these have worked for me. I seek your advice on which threshold value has proven effective for you or whether lower thresholding is safe in the competition. Your insights are greatly appreciated.
Thank you for your support.
", "messages": [ { "id": 2377997, "postDate": "2023-08-07T12:28:29.157Z", "content": "Hi all!
\nAs we approach the competition conclusion, Everyone are attempting to move up the leaderboard as the game draws to a close. To determine the ideal threshold value, I am now doing experiments. Lower thresholds are recommended by publicly accessible code and discussion. So I try 0.45 or 0.40, but neither of these have worked for me. I seek your advice on which threshold value has proven effective for you or whether lower thresholding is safe in the competition. Your insights are greatly appreciated.
Thank you for your support.
", "rawMarkdown": "Hi all!\nAs we approach the competition conclusion, Everyone are attempting to move up the leaderboard as the game draws to a close. To determine the ideal threshold value, I am now doing experiments. Lower thresholds are recommended by publicly accessible code and discussion. So I try 0.45 or 0.40, but neither of these have worked for me. I seek your advice on which threshold value has proven effective for you or whether lower thresholding is safe in the competition. Your insights are greatly appreciated.\n\nThank you for your support.", "votes": 3 }, { "id": 2379629, "postDate": "2023-08-08T08:31:34.047Z", "content": "@vipin20 The thing that I saw with thresholds is that in many cases models output almost all probabilities to 0.0 or to 1.0 and very few in between. Then, in these cases, threshold has a very low impact and also this explains why some people are even going down to 0.01 or 0.02 thresholds. Here you can see a couple of random examples of one of my models predicting on the val set and the probs distribution (y log scaled):
\n
you can see in the first example that even when it is wrong, it is predicting the pixels of that false positive contrail so hard to 1.0 (then I should also check if that was a mislabeled case and the model is actually right or if it's simply wrong)
", "rawMarkdown": "you can see in the first example that even when it is wrong, it is predicting the pixels of that false positive contrail so hard to 1.0 (then I should also check if that was a mislabeled case and the model is actually right or if it's simply wrong)", "replies": [ { "id": 2379661, "postDate": "2023-08-08T08:47:33.360Z", "content": "Thanks @edomingo Now i understand why people using very lower THR
", "rawMarkdown": "Thanks @edomingo Now i understand why people using very lower THR\n" } ] } ] } ], "comments": [ { "id": 2379629, "author_name": "Enric Domingo", "author_url": "", "post_date": "2023-08-08T08:31:34.047000", "content": "@vipin20 The thing that I saw with thresholds is that in many cases models output almost all probabilities to 0.0 or to 1.0 and very few in between. Then, in these cases, threshold has a very low impact and also this explains why some people are even going down to 0.01 or 0.02 thresholds. Here you can see a couple of random examples of one of my models predicting on the val set and the probs distribution (y log scaled):
\n
you can see in the first example that even when it is wrong, it is predicting the pixels of that false positive contrail so hard to 1.0 (then I should also check if that was a mislabeled case and the model is actually right or if it's simply wrong)
", "votes": 0, "replies": [ { "id": 2379661, "author_name": "Vipin Kumar", "author_url": "", "post_date": "2023-08-08T08:47:33.360000", "content": "Thanks @edomingo Now i understand why people using very lower THR
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2377997": "Hi all!\nAs we approach the competition conclusion, Everyone are attempting to move up the leaderboard as the game draws to a close. To determine the ideal threshold value, I am now doing experiments. Lower thresholds are recommended by publicly accessible code and discussion. So I try 0.45 or 0.40, but neither of these have worked for me. I seek your advice on which threshold value has proven effective for you or whether lower thresholding is safe in the competition. Your insights are greatly appreciated.\n\nThank you for your support.", "2379629": "@vipin20 The thing that I saw with thresholds is that in many cases models output almost all probabilities to 0.0 or to 1.0 and very few in between. Then, in these cases, threshold has a very low impact and also this explains why some people are even going down to 0.01 or 0.02 thresholds. Here you can see a couple of random examples of one of my models predicting on the val set and the probs distribution (y log scaled): \n" } }