So far my best LB score is 0.66 using only the competition data, I haven't tried any old competitions data yet would update this post once I did, if anyone did could they share what's your best score with and without the previous competitions data?, this info might help the community to push the scores to a better end.
", "messages": [ { "id": 2757753, "postDate": "2024-04-17T17:40:38.537Z", "content": "I used pre-trained models from 2022 and 2023 data, score ~ 0.65; without using pre-trained models, the score ~ 0.64.
", "rawMarkdown": "I used pre-trained models from 2022 and 2023 data, score ~ 0.65; without using pre-trained models, the score ~ 0.64.", "votes": 7, "replies": [ { "id": 2758206, "postDate": "2024-04-18T02:51:12.337Z", "content": "Thanks for the sharing
", "rawMarkdown": "Thanks for the sharing " } ] }, { "id": 2757364, "postDate": "2024-04-17T13:09:15.197Z", "content": "So far my best LB score is 0.66 using only the competition data, I haven't tried any old competitions data yet would update this post once I did, if anyone did could they share what's your best score with and without the previous competitions data?, this info might help the community to push the scores to a better end.
", "rawMarkdown": "So far my best LB score is 0.66 using only the competition data, I haven't tried any old competitions data yet would update this post once I did, if anyone did could they share what's your best score with and without the previous competitions data?, this info might help the community to push the scores to a better end.", "votes": 5 }, { "id": 2757663, "postDate": "2024-04-17T16:43:14.563Z", "content": "My most stable model scores ~0.665 (5 is approx), and this same model trained without the external data scores ~0.655 (again the third digit is approx)
", "rawMarkdown": "My most stable model scores ~0.665 (5 is approx), and this same model trained without the external data scores ~0.655 (again the third digit is approx)", "votes": 2, "replies": [ { "id": 2757749, "postDate": "2024-04-17T17:38:16.910Z", "content": "Oh thanks for letting me know, did you start ensemby? I have seen you on previous BirdCLF leaderboard, the last years scores reached .8+ public score, do you think that could also be the case for this competition?
", "rawMarkdown": "Oh thanks for letting me know, did you start ensemby? I have seen you on previous BirdCLF leaderboard, the last years scores reached .8+ public score, do you think that could also be the case for this competition?" } ] }, { "id": 2757595, "postDate": "2024-04-17T15:52:58.023Z", "content": "I tried pre-training with the data from BirdCLEF 2021-2023, but the improvements were minimal in my case. I'm curious if you've already started ensembling or if it's a single model that got you 66% LB?
", "rawMarkdown": "I tried pre-training with the data from BirdCLEF 2021-2023, but the improvements were minimal in my case. I'm curious if you've already started ensembling or if it's a single model that got you 66% LB?", "votes": 2, "replies": [ { "id": 2757623, "postDate": "2024-04-17T16:22:29.610Z", "content": "I think the training data is already huge with the number of 5 sec samples we can make , it does adds a large diversity even without any external data, and no i did not use ensembly yet, I have checked previous competitions score and they reached public score of 0.8+ so I think we should atleast get 0.75+ single model score before ensembly
", "rawMarkdown": "I think the training data is already huge with the number of 5 sec samples we can make , it does adds a large diversity even without any external data, and no i did not use ensembly yet, I have checked previous competitions score and they reached public score of 0.8+ so I think we should atleast get 0.75+ single model score before ensembly", "replies": [ { "id": 2757690, "postDate": "2024-04-17T17:03:43.073Z", "content": "Are you also using the unlabeled data?
", "rawMarkdown": "Are you also using the unlabeled data?" }, { "id": 2757738, "postDate": "2024-04-17T17:34:27.640Z", "content": "no I am not using them as of now,
", "rawMarkdown": "no I am not using them as of now," }, { "id": 2757752, "postDate": "2024-04-17T17:40:04.767Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2757757, "postDate": "2024-04-17T17:41:34.383Z", "content": "I used 5kfolds and there was a lot of variation per fold. I wonder if not using your ensemble means you used 1 fold out of 5 or multiple folds out of 5. And I'm curious if you've seen score differentials by fold as well.
", "rawMarkdown": "I used 5kfolds and there was a lot of variation per fold. I wonder if not using your ensemble means you used 1 fold out of 5 or multiple folds out of 5. And I'm curious if you've seen score differentials by fold as well." }, { "id": 2757826, "postDate": "2024-04-17T18:20:08.423Z", "content": "I submitted 5 folds of the same model training, 1 fold scores 0.66 and other scores low 0.64s - low 0.65s, and this is the most stable one, The model I had 3 days into the competition gets 0.67 on a fold and 0.62 on another
", "rawMarkdown": "I submitted 5 folds of the same model training, 1 fold scores 0.66 and other scores low 0.64s - low 0.65s, and this is the most stable one, The model I had 3 days into the competition gets 0.67 on a fold and 0.62 on another", "votes": 3 }, { "id": 2757847, "postDate": "2024-04-17T18:44:19.893Z", "content": "Thank you for sharing this valuable information with us😃
", "rawMarkdown": "Thank you for sharing this valuable information with us😃" }, { "id": 2758205, "postDate": "2024-04-18T02:50:01.447Z", "content": "I am using a normal test and train Split, and yes I did find visible variation in the scores when using different Splits
", "rawMarkdown": "I am using a normal test and train Split, and yes I did find visible variation in the scores when using different Splits " } ] } ] } ], "comments": [ { "id": 2757753, "author_name": "", "author_url": "", "post_date": "2024-04-17T17:40:38.537000", "content": "I used pre-trained models from 2022 and 2023 data, score ~ 0.65; without using pre-trained models, the score ~ 0.64.
", "votes": 7, "replies": [ { "id": 2758206, "author_name": "Arun", "author_url": "", "post_date": "2024-04-18T02:51:12.337000", "content": "Thanks for the sharing
", "votes": 0, "replies": [] } ] }, { "id": 2757663, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2024-04-17T16:43:14.563000", "content": "My most stable model scores ~0.665 (5 is approx), and this same model trained without the external data scores ~0.655 (again the third digit is approx)
", "votes": 2, "replies": [ { "id": 2757749, "author_name": "Arun", "author_url": "", "post_date": "2024-04-17T17:38:16.910000", "content": "Oh thanks for letting me know, did you start ensemby? I have seen you on previous BirdCLF leaderboard, the last years scores reached .8+ public score, do you think that could also be the case for this competition?
", "votes": 0, "replies": [] } ] }, { "id": 2757595, "author_name": "Hicham Bellafkir", "author_url": "", "post_date": "2024-04-17T15:52:58.023000", "content": "I tried pre-training with the data from BirdCLEF 2021-2023, but the improvements were minimal in my case. I'm curious if you've already started ensembling or if it's a single model that got you 66% LB?
", "votes": 2, "replies": [ { "id": 2757623, "author_name": "Arun", "author_url": "", "post_date": "2024-04-17T16:22:29.610000", "content": "I think the training data is already huge with the number of 5 sec samples we can make , it does adds a large diversity even without any external data, and no i did not use ensembly yet, I have checked previous competitions score and they reached public score of 0.8+ so I think we should atleast get 0.75+ single model score before ensembly
", "votes": 0, "replies": [ { "id": 2757690, "author_name": "Hicham Bellafkir", "author_url": "", "post_date": "2024-04-17T17:03:43.073000", "content": "Are you also using the unlabeled data?
", "votes": 0, "replies": [] }, { "id": 2757738, "author_name": "Arun", "author_url": "", "post_date": "2024-04-17T17:34:27.640000", "content": "no I am not using them as of now,
", "votes": 0, "replies": [] }, { "id": 2757752, "author_name": "", "author_url": "", "post_date": "2024-04-17T17:40:04.767000", "content": "", "votes": 0, "replies": [] }, { "id": 2757757, "author_name": "HB", "author_url": "", "post_date": "2024-04-17T17:41:34.383000", "content": "I used 5kfolds and there was a lot of variation per fold. I wonder if not using your ensemble means you used 1 fold out of 5 or multiple folds out of 5. And I'm curious if you've seen score differentials by fold as well.
", "votes": 0, "replies": [] }, { "id": 2757826, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2024-04-17T18:20:08.423000", "content": "I submitted 5 folds of the same model training, 1 fold scores 0.66 and other scores low 0.64s - low 0.65s, and this is the most stable one, The model I had 3 days into the competition gets 0.67 on a fold and 0.62 on another
", "votes": 3, "replies": [] }, { "id": 2757847, "author_name": "HB", "author_url": "", "post_date": "2024-04-17T18:44:19.893000", "content": "Thank you for sharing this valuable information with us😃
", "votes": 0, "replies": [] }, { "id": 2758205, "author_name": "Arun", "author_url": "", "post_date": "2024-04-18T02:50:01.447000", "content": "I am using a normal test and train Split, and yes I did find visible variation in the scores when using different Splits
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2757753": "I used pre-trained models from 2022 and 2023 data, score ~ 0.65; without using pre-trained models, the score ~ 0.64.", "2757364": "So far my best LB score is 0.66 using only the competition data, I haven't tried any old competitions data yet would update this post once I did, if anyone did could they share what's your best score with and without the previous competitions data?, this info might help the community to push the scores to a better end.", "2757663": "My most stable model scores ~0.665 (5 is approx), and this same model trained without the external data scores ~0.655 (again the third digit is approx)", "2757595": "I tried pre-training with the data from BirdCLEF 2021-2023, but the improvements were minimal in my case. I'm curious if you've already started ensembling or if it's a single model that got you 66% LB?" } }