Helo,
\nI am new to compeittion and I am a bit stuck. My CV and LB is consistent.
\nWhen I do 5-fold my score decreases. I am not sure what is happening. I want to understand if anyone got 0.5+ with a single model here?
", "messages": [ { "id": "1743209", "postDate": "04/02/2022 18:24:28", "content": "Helo,
\nI am new to compeittion and I am a bit stuck. My CV and LB is consistent.
\nWhen I do 5-fold my score decreases. I am not sure what is happening. I want to understand if anyone got 0.5+ with a single model here?
", "rawMarkdown": "Helo,\n\nI am new to compeittion and I am a bit stuck. My CV and LB is consistent.\n\nWhen I do 5-fold my score decreases. I am not sure what is happening. I want to understand if anyone got 0.5+ with a single model here?", "votes": null }, { "id": "1743294", "postDate": "04/02/2022 20:35:02", "content": "Yes. I can get a score > 0.5 with a single model
", "rawMarkdown": "Yes. I can get a score > 0.5 with a single model", "votes": null }, { "id": "1743298", "postDate": "04/02/2022 20:40:23", "content": "Can you please suggest changes I should do to improve my score, I am a student I don't want to win I just want to learn.
", "rawMarkdown": "Can you please suggest changes I should do to improve my score, I am a student I don't want to win I just want to learn.", "votes": null }, { "id": "1743301", "postDate": "04/02/2022 20:52:59", "content": "I tried using VGG19, Resnet50, EfficientNetB0 as my models on the spectrogram just like the public notebook of yours. But I am never able to cross 0.46.
", "rawMarkdown": "I tried using VGG19, Resnet50, EfficientNetB0 as my models on the spectrogram just like the public notebook of yours. But I am never able to cross 0.46.", "votes": null }, { "id": "1743302", "postDate": "04/02/2022 20:53:25", "content": "Do you have some base notebook that I can suggest changes? For me personally, an image classification model (for example a resnext50) trainning on spectrogram can give you > 0.5 score. You can take a look at my nb here: https://www.kaggle.com/code/dienhoa/fastai-pogchamp-music-genre-classification . I think just training a little bit longer can give you a better score
", "rawMarkdown": "Do you have some base notebook that I can suggest changes? For me personally, an image classification model (for example a resnext50) trainning on spectrogram can give you > 0.5 score. You can take a look at my nb here: https://www.kaggle.com/code/dienhoa/fastai-pogchamp-music-genre-classification . I think just training a little bit longer can give you a better score", "votes": null }, { "id": "1743304", "postDate": "04/02/2022 20:55:38", "content": "What is your score in CV ? In that case, you can try some augmentation to make your model more robust on the LB, mixup, cropping your audio, changing the loudness, …
", "rawMarkdown": "What is your score in CV ? In that case, you can try some augmentation to make your model more robust on the LB, mixup, cropping your audio, changing the loudness, ...", "votes": null }, { "id": "1743307", "postDate": "04/02/2022 20:58:44", "content": "Thank you so much for your suggestion of ResNext50. I am following @harveenchadha baseline notebook as I can understand keras much better but I am not sure what all I can tune in that notebook.
\nhttps://www.kaggle.com/code/harveenchadha/pog-1-spect-2d-cnn-keras-baseline
", "rawMarkdown": "Thank you so much for your suggestion of ResNext50. I am following @harveenchadha baseline notebook as I can understand keras much better but I am not sure what all I can tune in that notebook.\n\nhttps://www.kaggle.com/code/harveenchadha/pog-1-spect-2d-cnn-keras-baseline", "votes": null }, { "id": "1743308", "postDate": "04/02/2022 21:00:42", "content": "My CV is in line with LB I think. My accuracy is nearly equal to CV score. I am not able to understand how CV accuracy and LB f-1 score are correlated either.
", "rawMarkdown": "My CV is in line with LB I think. My accuracy is nearly equal to CV score. I am not able to understand how CV accuracy and LB f-1 score are correlated either.", "votes": null }, { "id": "1743309", "postDate": "04/02/2022 21:00:50", "content": "Let's try what I've just said. I'm sure in keras you can do all of these stuffs
", "rawMarkdown": "Let's try what I've just said. I'm sure in keras you can do all of these stuffs", "votes": null }, { "id": "1743310", "postDate": "04/02/2022 21:03:31", "content": "For single label classification, I think f1-micro is equivalent to accuracy https://stackoverflow.com/questions/37358496/is-f1-micro-the-same-as-accuracy#:~:text=4%20Answers&text=is%20not%20useful-,Show%20activity%20on%20this%20post.,case%20in%20multi-label%20classification
", "rawMarkdown": "For single label classification, I think f1-micro is equivalent to accuracy https://stackoverflow.com/questions/37358496/is-f1-micro-the-same-as-accuracy#:~:text=4%20Answers&text=is%20not%20useful-,Show%20activity%20on%20this%20post.,case%20in%20multi-label%20classification", "votes": null }, { "id": "1743311", "postDate": "04/02/2022 21:05:05", "content": "Thank you, i will keep on experimenting and will post the results here. the LB calculation metric is not macro F-1?
", "rawMarkdown": "Thank you, i will keep on experimenting and will post the results here. the LB calculation metric is not macro F-1?", "votes": null }, { "id": "1743313", "postDate": "04/02/2022 21:07:01", "content": "No, it's f1 micro: https://www.kaggle.com/competitions/kaggle-pog-series-s01e02/discussion/313286
", "rawMarkdown": "No, it's f1 micro: https://www.kaggle.com/competitions/kaggle-pog-series-s01e02/discussion/313286", "votes": null }, { "id": "1743315", "postDate": "04/02/2022 21:08:07", "content": "Thank you so much, all this time I was focusing on macro-F1. Thanks again for your help!
", "rawMarkdown": "Thank you so much, all this time I was focusing on macro-F1. Thanks again for your help!", "votes": null }, { "id": "1743321", "postDate": "04/02/2022 21:13:27", "content": "I'm glad to help :)
", "rawMarkdown": "I'm glad to help :)", "votes": null }, { "id": "1744328", "postDate": "04/03/2022 21:39:18", "content": "I just tried to replace custom architecture with effnet in my public notebook and scored 0.50+ , so in short it is possible.
", "rawMarkdown": "I just tried to replace custom architecture with effnet in my public notebook and scored 0.50+ , so in short it is possible.", "votes": null } ], "comments": [ { "id": 1743294, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 20:35:02", "content": "Yes. I can get a score > 0.5 with a single model
", "votes": null, "replies": [ { "id": 1743298, "author_name": "rahultiwari1995", "author_url": "", "post_date": "04/02/2022 20:40:23", "content": "Can you please suggest changes I should do to improve my score, I am a student I don't want to win I just want to learn.
", "votes": null, "replies": [] }, { "id": 1743301, "author_name": "rahultiwari1995", "author_url": "", "post_date": "04/02/2022 20:52:59", "content": "I tried using VGG19, Resnet50, EfficientNetB0 as my models on the spectrogram just like the public notebook of yours. But I am never able to cross 0.46.
", "votes": null, "replies": [] }, { "id": 1743302, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 20:53:25", "content": "Do you have some base notebook that I can suggest changes? For me personally, an image classification model (for example a resnext50) trainning on spectrogram can give you > 0.5 score. You can take a look at my nb here: https://www.kaggle.com/code/dienhoa/fastai-pogchamp-music-genre-classification . I think just training a little bit longer can give you a better score
", "votes": null, "replies": [] }, { "id": 1743304, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 20:55:38", "content": "What is your score in CV ? In that case, you can try some augmentation to make your model more robust on the LB, mixup, cropping your audio, changing the loudness, …
", "votes": null, "replies": [] }, { "id": 1743307, "author_name": "rahultiwari1995", "author_url": "", "post_date": "04/02/2022 20:58:44", "content": "Thank you so much for your suggestion of ResNext50. I am following @harveenchadha baseline notebook as I can understand keras much better but I am not sure what all I can tune in that notebook.
\nhttps://www.kaggle.com/code/harveenchadha/pog-1-spect-2d-cnn-keras-baseline
", "votes": null, "replies": [] }, { "id": 1743308, "author_name": "rahultiwari1995", "author_url": "", "post_date": "04/02/2022 21:00:42", "content": "My CV is in line with LB I think. My accuracy is nearly equal to CV score. I am not able to understand how CV accuracy and LB f-1 score are correlated either.
", "votes": null, "replies": [] }, { "id": 1743309, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 21:00:50", "content": "Let's try what I've just said. I'm sure in keras you can do all of these stuffs
", "votes": null, "replies": [] }, { "id": 1743310, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 21:03:31", "content": "For single label classification, I think f1-micro is equivalent to accuracy https://stackoverflow.com/questions/37358496/is-f1-micro-the-same-as-accuracy#:~:text=4%20Answers&text=is%20not%20useful-,Show%20activity%20on%20this%20post.,case%20in%20multi-label%20classification
", "votes": null, "replies": [] }, { "id": 1743311, "author_name": "rahultiwari1995", "author_url": "", "post_date": "04/02/2022 21:05:05", "content": "Thank you, i will keep on experimenting and will post the results here. the LB calculation metric is not macro F-1?
", "votes": null, "replies": [] }, { "id": 1743313, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 21:07:01", "content": "No, it's f1 micro: https://www.kaggle.com/competitions/kaggle-pog-series-s01e02/discussion/313286
", "votes": null, "replies": [] }, { "id": 1743315, "author_name": "rahultiwari1995", "author_url": "", "post_date": "04/02/2022 21:08:07", "content": "Thank you so much, all this time I was focusing on macro-F1. Thanks again for your help!
", "votes": null, "replies": [] }, { "id": 1743321, "author_name": "dienhoa", "author_url": "", "post_date": "04/02/2022 21:13:27", "content": "I'm glad to help :)
", "votes": null, "replies": [] } ] }, { "id": 1744328, "author_name": "harveenchadha", "author_url": "", "post_date": "04/03/2022 21:39:18", "content": "I just tried to replace custom architecture with effnet in my public notebook and scored 0.50+ , so in short it is possible.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1743209": "Helo,\n\nI am new to compeittion and I am a bit stuck. My CV and LB is consistent.\n\nWhen I do 5-fold my score decreases. I am not sure what is happening. I want to understand if anyone got 0.5+ with a single model here?", "1743294": "Yes. I can get a score > 0.5 with a single model", "1743298": "Can you please suggest changes I should do to improve my score, I am a student I don't want to win I just want to learn.", "1743301": "I tried using VGG19, Resnet50, EfficientNetB0 as my models on the spectrogram just like the public notebook of yours. But I am never able to cross 0.46.", "1743302": "Do you have some base notebook that I can suggest changes? For me personally, an image classification model (for example a resnext50) trainning on spectrogram can give you > 0.5 score. You can take a look at my nb here: https://www.kaggle.com/code/dienhoa/fastai-pogchamp-music-genre-classification . I think just training a little bit longer can give you a better score", "1743304": "What is your score in CV ? In that case, you can try some augmentation to make your model more robust on the LB, mixup, cropping your audio, changing the loudness, ...", "1743307": "Thank you so much for your suggestion of ResNext50. I am following @harveenchadha baseline notebook as I can understand keras much better but I am not sure what all I can tune in that notebook.\n\nhttps://www.kaggle.com/code/harveenchadha/pog-1-spect-2d-cnn-keras-baseline", "1743308": "My CV is in line with LB I think. My accuracy is nearly equal to CV score. I am not able to understand how CV accuracy and LB f-1 score are correlated either.", "1743309": "Let's try what I've just said. I'm sure in keras you can do all of these stuffs", "1743310": "For single label classification, I think f1-micro is equivalent to accuracy https://stackoverflow.com/questions/37358496/is-f1-micro-the-same-as-accuracy#:~:text=4%20Answers&text=is%20not%20useful-,Show%20activity%20on%20this%20post.,case%20in%20multi-label%20classification", "1743311": "Thank you, i will keep on experimenting and will post the results here. the LB calculation metric is not macro F-1?", "1743313": "No, it's f1 micro: https://www.kaggle.com/competitions/kaggle-pog-series-s01e02/discussion/313286", "1743315": "Thank you so much, all this time I was focusing on macro-F1. Thanks again for your help!", "1743321": "I'm glad to help :)", "1744328": "I just tried to replace custom architecture with effnet in my public notebook and scored 0.50+ , so in short it is possible." }, "source": "meta" }