I am trying to understand this competition more and was wondering what will and won't help my score. Thank you so much for your attention.
", "messages": [ { "id": 2811854, "postDate": "2024-05-14T00:39:50.353Z", "content": "The list is too long to share here.
", "rawMarkdown": "The list is too long to share here.", "votes": 6, "replies": [ { "id": 2811879, "postDate": "2024-05-14T01:18:00.233Z", "content": "Any main things that were helpful in previous competitions and not this one? If you do not want to share any secrets about your solution to the problem that's fine. Keep up the good work in the first place by the way I'm waiting for an interesting notebook reveal or discussion at the end of this competition😉!
", "rawMarkdown": "Any main things that were helpful in previous competitions and not this one? If you do not want to share any secrets about your solution to the problem that's fine. Keep up the good work in the first place by the way I'm waiting for an interesting notebook reveal or discussion at the end of this competition😉!", "replies": [ { "id": 2811982, "postDate": "2024-05-14T03:48:06.333Z", "content": "@cpmpml all too relatable, my first thought was “well mostly everything” lol still plenty of time left but very curious to see the leading solutions on this one.
", "rawMarkdown": "@cpmpml all too relatable, my first thought was “well mostly everything” lol still plenty of time left but very curious to see the leading solutions on this one. ", "votes": 1 }, { "id": 2812543, "postDate": "2024-05-14T09:22:58.287Z", "content": "In this year competition you cannot gain an advantage by gaming the metric (like post processing predictions class by class). This year competition is much better at evaluating model quality IMHO.
\nGiven how kaggle community is good at adapting to competition metric, maybe quite a lot of past winning solutions just isn't useful here.
\nAlso, this year we have 2.2x more data to score in the same time. It means that none of the past top solutions would run fast enough here.
", "rawMarkdown": "In this year competition you cannot gain an advantage by gaming the metric (like post processing predictions class by class). This year competition is much better at evaluating model quality IMHO. \n\nGiven how kaggle community is good at adapting to competition metric, maybe quite a lot of past winning solutions just isn't useful here. \n\nAlso, this year we have 2.2x more data to score in the same time. It means that none of the past top solutions would run fast enough here. ", "votes": 3 }, { "id": 2813069, "postDate": "2024-05-14T14:51:00.810Z", "content": "I agree, this years seems to grab the challenges of the real world a bit better. My only wish is that the train and test data were a little more similar but still a good comp, just a tough one. Good learning experience will come from it
", "rawMarkdown": "I agree, this years seems to grab the challenges of the real world a bit better. My only wish is that the train and test data were a little more similar but still a good comp, just a tough one. Good learning experience will come from it ", "votes": 1 }, { "id": 2813107, "postDate": "2024-05-14T15:18:55.950Z", "content": "Yes the data in train and test is very different I believe that CPMP made a good comparison here https://www.kaggle.com/competitions/birdclef-2024/discussion/498404
", "rawMarkdown": "Yes the data in train and test is very different I believe that CPMP made a good comparison here https://www.kaggle.com/competitions/birdclef-2024/discussion/498404", "votes": 2 } ] } ] }, { "id": 2821203, "postDate": "2024-05-18T02:10:49.103Z", "content": "Encodec language model. It trained well, but was far too slow for inference. I haven't tried to distill it though.
", "rawMarkdown": "Encodec language model. It trained well, but was far too slow for inference. I haven't tried to distill it though.", "votes": 1, "replies": [ { "id": 2822257, "postDate": "2024-05-18T14:07:49.567Z", "content": "Yes KD could prove useful as it has been used in many of the previous years best works!
", "rawMarkdown": "Yes KD could prove useful as it has been used in many of the previous years best works!", "votes": 1 } ] }, { "id": 2814663, "postDate": "2024-05-15T13:31:19.477Z", "content": "Related thread: https://www.kaggle.com/competitions/birdclef-2024/discussion/502401
", "rawMarkdown": "Related thread: https://www.kaggle.com/competitions/birdclef-2024/discussion/502401", "votes": 1, "replies": [ { "id": 2814887, "postDate": "2024-05-15T15:45:07.197Z", "content": "Interesting thank you!
", "rawMarkdown": "Interesting thank you!" } ] }, { "id": 2811899, "postDate": "2024-05-14T01:46:13.610Z", "content": "I tried some methods to solve the problem of class imbalance, such as oversampling and loss weighting, but none of them were successful. 😭
", "rawMarkdown": "I tried some methods to solve the problem of class imbalance, such as oversampling and loss weighting, but none of them were successful. 😭", "votes": 1, "replies": [ { "id": 2811929, "postDate": "2024-05-14T02:19:59.653Z", "content": "Yes, I also tried the methods you suggested and my LB score decreased quite a bit. This competition has some weird data. 🤔
", "rawMarkdown": "Yes, I also tried the methods you suggested and my LB score decreased quite a bit. This competition has some weird data. 🤔" } ] }, { "id": 2811852, "postDate": "2024-05-14T00:38:25.447Z", "content": "I am trying to understand this competition more and was wondering what will and won't help my score. Thank you so much for your attention.
", "rawMarkdown": "I am trying to understand this competition more and was wondering what will and won't help my score. Thank you so much for your attention.", "votes": 1 }, { "id": 2814830, "postDate": "2024-05-15T15:16:41.407Z", "content": "For me training the model on all 5 sec segments gave much worse LB than training of first 5 sec only.
", "rawMarkdown": "For me training the model on all 5 sec segments gave much worse LB than training of first 5 sec only.", "votes": 2, "replies": [ { "id": 2814883, "postDate": "2024-05-15T15:44:26.453Z", "content": "Interesting because I have read some reports of people's cores drastically increasing! Out of luck, you caught me right when I was trying this strategy so I can test if it helps me or not. My CV is currently really good compared to previous tries!
", "rawMarkdown": "Interesting because I have read some reports of people's cores drastically increasing! Out of luck, you caught me right when I was trying this strategy so I can test if it helps me or not. My CV is currently really good compared to previous tries!", "replies": [ { "id": 2814916, "postDate": "2024-05-15T16:00:58.040Z", "content": "CV is really high, but LB is not so good. I has the same results when I trained model on all available Xeno data for scored classes.
", "rawMarkdown": "CV is really high, but LB is not so good. I has the same results when I trained model on all available Xeno data for scored classes.", "votes": 1 }, { "id": 2814935, "postDate": "2024-05-15T16:07:41.500Z", "content": "Interesting my model might be different or something because when using xeno canto it made my model better on CV and LB
", "rawMarkdown": "Interesting my model might be different or something because when using xeno canto it made my model better on CV and LB" } ] } ] }, { "id": 2814741, "postDate": "2024-05-15T14:20:55.840Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2811854, "author_name": "CPMP", "author_url": "", "post_date": "2024-05-14T00:39:50.353000", "content": "The list is too long to share here.
", "votes": 6, "replies": [ { "id": 2811879, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-14T01:18:00.233000", "content": "Any main things that were helpful in previous competitions and not this one? If you do not want to share any secrets about your solution to the problem that's fine. Keep up the good work in the first place by the way I'm waiting for an interesting notebook reveal or discussion at the end of this competition😉!
", "votes": 0, "replies": [ { "id": 2811982, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-05-14T03:48:06.333000", "content": "@cpmpml all too relatable, my first thought was “well mostly everything” lol still plenty of time left but very curious to see the leading solutions on this one.
", "votes": 1, "replies": [] }, { "id": 2812543, "author_name": "CPMP", "author_url": "", "post_date": "2024-05-14T09:22:58.287000", "content": "In this year competition you cannot gain an advantage by gaming the metric (like post processing predictions class by class). This year competition is much better at evaluating model quality IMHO.
\nGiven how kaggle community is good at adapting to competition metric, maybe quite a lot of past winning solutions just isn't useful here.
\nAlso, this year we have 2.2x more data to score in the same time. It means that none of the past top solutions would run fast enough here.
", "votes": 3, "replies": [] }, { "id": 2813069, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-05-14T14:51:00.810000", "content": "I agree, this years seems to grab the challenges of the real world a bit better. My only wish is that the train and test data were a little more similar but still a good comp, just a tough one. Good learning experience will come from it
", "votes": 1, "replies": [] }, { "id": 2813107, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-14T15:18:55.950000", "content": "Yes the data in train and test is very different I believe that CPMP made a good comparison here https://www.kaggle.com/competitions/birdclef-2024/discussion/498404
", "votes": 2, "replies": [] } ] } ] }, { "id": 2821203, "author_name": "Will Rice", "author_url": "", "post_date": "2024-05-18T02:10:49.103000", "content": "Encodec language model. It trained well, but was far too slow for inference. I haven't tried to distill it though.
", "votes": 1, "replies": [ { "id": 2822257, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-18T14:07:49.567000", "content": "Yes KD could prove useful as it has been used in many of the previous years best works!
", "votes": 1, "replies": [] } ] }, { "id": 2814663, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2024-05-15T13:31:19.477000", "content": "Related thread: https://www.kaggle.com/competitions/birdclef-2024/discussion/502401
", "votes": 1, "replies": [ { "id": 2814887, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-15T15:45:07.197000", "content": "Interesting thank you!
", "votes": 0, "replies": [] } ] }, { "id": 2811899, "author_name": "zznznb", "author_url": "", "post_date": "2024-05-14T01:46:13.610000", "content": "I tried some methods to solve the problem of class imbalance, such as oversampling and loss weighting, but none of them were successful. 😭
", "votes": 1, "replies": [ { "id": 2811929, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-14T02:19:59.653000", "content": "Yes, I also tried the methods you suggested and my LB score decreased quite a bit. This competition has some weird data. 🤔
", "votes": 0, "replies": [] } ] }, { "id": 2814830, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2024-05-15T15:16:41.407000", "content": "For me training the model on all 5 sec segments gave much worse LB than training of first 5 sec only.
", "votes": 2, "replies": [ { "id": 2814883, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-15T15:44:26.453000", "content": "Interesting because I have read some reports of people's cores drastically increasing! Out of luck, you caught me right when I was trying this strategy so I can test if it helps me or not. My CV is currently really good compared to previous tries!
", "votes": 0, "replies": [ { "id": 2814916, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2024-05-15T16:00:58.040000", "content": "CV is really high, but LB is not so good. I has the same results when I trained model on all available Xeno data for scored classes.
", "votes": 1, "replies": [] }, { "id": 2814935, "author_name": "MaxG6", "author_url": "", "post_date": "2024-05-15T16:07:41.500000", "content": "Interesting my model might be different or something because when using xeno canto it made my model better on CV and LB
", "votes": 0, "replies": [] } ] } ] }, { "id": 2814741, "author_name": "", "author_url": "", "post_date": "2024-05-15T14:20:55.840000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2811854": "The list is too long to share here.", "2821203": "Encodec language model. It trained well, but was far too slow for inference. I haven't tried to distill it though.", "2814663": "Related thread: https://www.kaggle.com/competitions/birdclef-2024/discussion/502401", "2811899": "I tried some methods to solve the problem of class imbalance, such as oversampling and loss weighting, but none of them were successful. 😭", "2811852": "I am trying to understand this competition more and was wondering what will and won't help my score. Thank you so much for your attention.", "2814830": "For me training the model on all 5 sec segments gave much worse LB than training of first 5 sec only.", "2814741": "" } }