Hello,
\n\nI released on a github repository my solution. It is highly inspired from @daisukelab notebooks for preprocessing and @mhiro2 for his simple CNN model.
\n\nKeys points of this solution, 2 techniques I imagined (maybe they exist under another name 😄) \n- warm-up pipeline : try to use noisy dataset as pretrain model and as semi-supervised to increase diversity in generating a set of models. In another words, first use noisy set to warmup model training, then fine tune model with curated set and then use noisy set again in a semi-supervised way.\n- SpecMix : my new data-augmentation technique which takes what I consider the best from SpecAugment and mixup. It generates new samples by applying frequency replacement and time replacement on inputs and compute a weighted average on targets. More details in the github README
\n\nInference is done using an ensemble of 2 models (@mhiro2 simple CNN model + VGG16) on 10 folds CV.
\n\nMore details are available in the github README, as well code and weights.
", "messages": [ { "id": "550652", "postDate": "06/11/2019 21:59:21", "content": "Hello,
\n\nI released on a github repository my solution. It is highly inspired from @daisukelab notebooks for preprocessing and @mhiro2 for his simple CNN model.
\n\nKeys points of this solution, 2 techniques I imagined (maybe they exist under another name 😄) \n- warm-up pipeline : try to use noisy dataset as pretrain model and as semi-supervised to increase diversity in generating a set of models. In another words, first use noisy set to warmup model training, then fine tune model with curated set and then use noisy set again in a semi-supervised way.\n- SpecMix : my new data-augmentation technique which takes what I consider the best from SpecAugment and mixup. It generates new samples by applying frequency replacement and time replacement on inputs and compute a weighted average on targets. More details in the github README
\n\nInference is done using an ensemble of 2 models (@mhiro2 simple CNN model + VGG16) on 10 folds CV.
\n\nMore details are available in the github README, as well code and weights.
", "rawMarkdown": "Hello,\n\nI released on a [github repository](https://github.com/ebouteillon/freesound-audio-tagging-2019) my solution. It is highly inspired from @daisukelab notebooks for preprocessing and @mhiro2 for his simple CNN model.\n\nKeys points of this solution, 2 techniques I imagined (maybe they exist under another name 😄) \n- **warm-up pipeline** : try to use noisy dataset as pretrain model and as semi-supervised to increase diversity in generating a set of models. In another words, first use noisy set to warmup model training, then fine tune model with curated set and then use noisy set again in a semi-supervised way.\n- **SpecMix** : my new data-augmentation technique which takes what I consider the best from SpecAugment and mixup. It generates new samples by applying frequency replacement and time replacement on inputs and compute a weighted average on targets. More details in the github [README](https://github.com/ebouteillon/freesound-audio-tagging-2019)\n\nInference is done using an ensemble of 2 models (@mhiro2 simple CNN model + VGG16) on 10 folds CV.\n\nMore details are available in the github [README](https://github.com/ebouteillon/freesound-audio-tagging-2019), as well code and weights.", "votes": null }, { "id": "550767", "postDate": "06/12/2019 02:26:31", "content": "Thank for sharing a solid solution!
", "rawMarkdown": "Thank for sharing a solid solution!", "votes": null }, { "id": "550789", "postDate": "06/12/2019 03:14:01", "content": "Thank you for sharing! Your solution is quite simple enough and strong.
", "rawMarkdown": "Thank you for sharing! Your solution is quite simple enough and strong.", "votes": null }, { "id": "550806", "postDate": "06/12/2019 03:36:28", "content": "Thank you for using many of my codes, it's my pleasure, really! # including it's easy for me to read your code ;)\nAnd your two ideas are really nice and making good performance actually, it's much better than straightforward mine.\nThis is exactly what I was expecting to see. Wishing your good luck in the 2nd stage!
", "rawMarkdown": "Thank you for using many of my codes, it's my pleasure, really! # including it's easy for me to read your code ;)\nAnd your two ideas are really nice and making good performance actually, it's much better than straightforward mine.\nThis is exactly what I was expecting to see. Wishing your good luck in the 2nd stage!", "votes": null }, { "id": "551146", "postDate": "06/12/2019 11:43:55", "content": "Thank you for your nice words.\nWishing you all the best for the second stage!
", "rawMarkdown": "Thank you for your nice words.\nWishing you all the best for the second stage!", "votes": null }, { "id": "551148", "postDate": "06/12/2019 11:46:52", "content": "It should be me that is thanking you for the model that I borrowed, it helped me a lot :smile:\nWishing you all the best for the second stage!
", "rawMarkdown": "It should be me that is thanking you for the model that I borrowed, it helped me a lot :smile:\nWishing you all the best for the second stage!", "votes": null }, { "id": "551149", "postDate": "06/12/2019 11:48:33", "content": "My code is a bit crappy, I should refactor it and comment it. Thanks for your kernels and nice comment. :smile:\nWishing you all the best for the second stage!
", "rawMarkdown": "My code is a bit crappy, I should refactor it and comment it. Thanks for your kernels and nice comment. :smile:\nWishing you all the best for the second stage!", "votes": null }, { "id": "551176", "postDate": "06/12/2019 12:29:45", "content": "thanks for sharing @ebouteillon .
", "rawMarkdown": "thanks for sharing @ebouteillon .", "votes": null }, { "id": "551364", "postDate": "06/12/2019 16:21:00", "content": "You are welcomed. Good luck!
", "rawMarkdown": "You are welcomed. Good luck!", "votes": null }, { "id": "551504", "postDate": "06/12/2019 19:46:49", "content": "This is fantastic and really well documented. Learned a lot, thanks for sharing.
", "rawMarkdown": "This is fantastic and really well documented. Learned a lot, thanks for sharing.", "votes": null }, { "id": "551613", "postDate": "06/12/2019 23:46:20", "content": "Your kind words are appreciated
", "rawMarkdown": "Your kind words are appreciated", "votes": null }, { "id": "551641", "postDate": "06/13/2019 00:56:27", "content": "Thanks for sharing @ebouteillon , warm-up pipeline really make sense.
", "rawMarkdown": "Thanks for sharing @ebouteillon , warm-up pipeline really make sense.", "votes": null }, { "id": "551845", "postDate": "06/13/2019 07:04:40", "content": "Thank you for your positive feedback. An important point to take care of is not leaking data between folds and stages during the warm-up.\nGood luck for the second stage.
", "rawMarkdown": "Thank you for your positive feedback. An important point to take care of is not leaking data between folds and stages during the warm-up.\nGood luck for the second stage.", "votes": null }, { "id": "553546", "postDate": "06/15/2019 22:14:21", "content": "Hi my Kaggling friends,\nJust to let you know that I updated the README in the repository. I tried to clarify some parts of the description, added instructions to reproduce my results, acknowledgments where due, and of course useless badges... 😄 \nFeedback are of course welcomed.
", "rawMarkdown": "Hi my Kaggling friends,\nJust to let you know that I updated the [README](https://github.com/ebouteillon/freesound-audio-tagging-2019) in the repository. I tried to clarify some parts of the description, added instructions to reproduce my results, acknowledgments where due, and of course useless badges... 😄 \nFeedback are of course welcomed.", "votes": null }, { "id": "557622", "postDate": "06/21/2019 10:24:38", "content": "“Dad, you are the best and you will be at the very top” . THIS is real motivation !!! Congrats :)
", "rawMarkdown": "“Dad, you are the best and you will be at the very top” . THIS is real motivation !!! Congrats :)", "votes": null }, { "id": "557659", "postDate": "06/21/2019 11:18:04", "content": "Thank for noticing it. All credits goes to him. 😉
", "rawMarkdown": "Thank for noticing it. All credits goes to him. 😉", "votes": null }, { "id": "563922", "postDate": "06/28/2019 19:28:59", "content": "Note: Updated title and happy to become master 🥳
", "rawMarkdown": "Note: Updated title and happy to become master 🥳", "votes": null }, { "id": "573149", "postDate": "07/11/2019 21:10:35", "content": "thank you so much for sharing
", "rawMarkdown": "thank you so much for sharing", "votes": null }, { "id": "573861", "postDate": "07/12/2019 21:07:14", "content": "You’re welcomed.
", "rawMarkdown": "You’re welcomed.", "votes": null } ], "comments": [ { "id": 550767, "author_name": "backaggle", "author_url": "", "post_date": "06/12/2019 02:26:31", "content": "Thank for sharing a solid solution!
", "votes": null, "replies": [ { "id": 551146, "author_name": "ebouteillon", "author_url": "", "post_date": "06/12/2019 11:43:55", "content": "Thank you for your nice words.\nWishing you all the best for the second stage!
", "votes": null, "replies": [] } ] }, { "id": 550789, "author_name": "mhiro2", "author_url": "", "post_date": "06/12/2019 03:14:01", "content": "Thank you for sharing! Your solution is quite simple enough and strong.
", "votes": null, "replies": [ { "id": 551148, "author_name": "ebouteillon", "author_url": "", "post_date": "06/12/2019 11:46:52", "content": "It should be me that is thanking you for the model that I borrowed, it helped me a lot :smile:\nWishing you all the best for the second stage!
", "votes": null, "replies": [] } ] }, { "id": 550806, "author_name": "daisukelab", "author_url": "", "post_date": "06/12/2019 03:36:28", "content": "Thank you for using many of my codes, it's my pleasure, really! # including it's easy for me to read your code ;)\nAnd your two ideas are really nice and making good performance actually, it's much better than straightforward mine.\nThis is exactly what I was expecting to see. Wishing your good luck in the 2nd stage!
", "votes": null, "replies": [ { "id": 551149, "author_name": "ebouteillon", "author_url": "", "post_date": "06/12/2019 11:48:33", "content": "My code is a bit crappy, I should refactor it and comment it. Thanks for your kernels and nice comment. :smile:\nWishing you all the best for the second stage!
", "votes": null, "replies": [] } ] }, { "id": 551176, "author_name": "karanjakhar", "author_url": "", "post_date": "06/12/2019 12:29:45", "content": "thanks for sharing @ebouteillon .
", "votes": null, "replies": [ { "id": 551364, "author_name": "ebouteillon", "author_url": "", "post_date": "06/12/2019 16:21:00", "content": "You are welcomed. Good luck!
", "votes": null, "replies": [] } ] }, { "id": 551504, "author_name": "madeupmasters", "author_url": "", "post_date": "06/12/2019 19:46:49", "content": "This is fantastic and really well documented. Learned a lot, thanks for sharing.
", "votes": null, "replies": [ { "id": 551613, "author_name": "ebouteillon", "author_url": "", "post_date": "06/12/2019 23:46:20", "content": "Your kind words are appreciated
", "votes": null, "replies": [] } ] }, { "id": 551641, "author_name": "sailorwei", "author_url": "", "post_date": "06/13/2019 00:56:27", "content": "Thanks for sharing @ebouteillon , warm-up pipeline really make sense.
", "votes": null, "replies": [ { "id": 551845, "author_name": "ebouteillon", "author_url": "", "post_date": "06/13/2019 07:04:40", "content": "Thank you for your positive feedback. An important point to take care of is not leaking data between folds and stages during the warm-up.\nGood luck for the second stage.
", "votes": null, "replies": [] } ] }, { "id": 553546, "author_name": "ebouteillon", "author_url": "", "post_date": "06/15/2019 22:14:21", "content": "Hi my Kaggling friends,\nJust to let you know that I updated the README in the repository. I tried to clarify some parts of the description, added instructions to reproduce my results, acknowledgments where due, and of course useless badges... 😄 \nFeedback are of course welcomed.
", "votes": null, "replies": [] }, { "id": 557622, "author_name": "nikosdim", "author_url": "", "post_date": "06/21/2019 10:24:38", "content": "“Dad, you are the best and you will be at the very top” . THIS is real motivation !!! Congrats :)
", "votes": null, "replies": [ { "id": 557659, "author_name": "ebouteillon", "author_url": "", "post_date": "06/21/2019 11:18:04", "content": "Thank for noticing it. All credits goes to him. 😉
", "votes": null, "replies": [] } ] }, { "id": 563922, "author_name": "ebouteillon", "author_url": "", "post_date": "06/28/2019 19:28:59", "content": "Note: Updated title and happy to become master 🥳
", "votes": null, "replies": [] }, { "id": 573149, "author_name": "lingzhikang", "author_url": "", "post_date": "07/11/2019 21:10:35", "content": "thank you so much for sharing
", "votes": null, "replies": [ { "id": 573861, "author_name": "ebouteillon", "author_url": "", "post_date": "07/12/2019 21:07:14", "content": "You’re welcomed.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "550652": "Hello,\n\nI released on a [github repository](https://github.com/ebouteillon/freesound-audio-tagging-2019) my solution. It is highly inspired from @daisukelab notebooks for preprocessing and @mhiro2 for his simple CNN model.\n\nKeys points of this solution, 2 techniques I imagined (maybe they exist under another name 😄) \n- **warm-up pipeline** : try to use noisy dataset as pretrain model and as semi-supervised to increase diversity in generating a set of models. In another words, first use noisy set to warmup model training, then fine tune model with curated set and then use noisy set again in a semi-supervised way.\n- **SpecMix** : my new data-augmentation technique which takes what I consider the best from SpecAugment and mixup. It generates new samples by applying frequency replacement and time replacement on inputs and compute a weighted average on targets. More details in the github [README](https://github.com/ebouteillon/freesound-audio-tagging-2019)\n\nInference is done using an ensemble of 2 models (@mhiro2 simple CNN model + VGG16) on 10 folds CV.\n\nMore details are available in the github [README](https://github.com/ebouteillon/freesound-audio-tagging-2019), as well code and weights.", "550767": "Thank for sharing a solid solution!", "550789": "Thank you for sharing! Your solution is quite simple enough and strong.", "550806": "Thank you for using many of my codes, it's my pleasure, really! # including it's easy for me to read your code ;)\nAnd your two ideas are really nice and making good performance actually, it's much better than straightforward mine.\nThis is exactly what I was expecting to see. Wishing your good luck in the 2nd stage!", "551146": "Thank you for your nice words.\nWishing you all the best for the second stage!", "551148": "It should be me that is thanking you for the model that I borrowed, it helped me a lot :smile:\nWishing you all the best for the second stage!", "551149": "My code is a bit crappy, I should refactor it and comment it. Thanks for your kernels and nice comment. :smile:\nWishing you all the best for the second stage!", "551176": "thanks for sharing @ebouteillon .", "551364": "You are welcomed. Good luck!", "551504": "This is fantastic and really well documented. Learned a lot, thanks for sharing.", "551613": "Your kind words are appreciated", "551641": "Thanks for sharing @ebouteillon , warm-up pipeline really make sense.", "551845": "Thank you for your positive feedback. An important point to take care of is not leaking data between folds and stages during the warm-up.\nGood luck for the second stage.", "553546": "Hi my Kaggling friends,\nJust to let you know that I updated the [README](https://github.com/ebouteillon/freesound-audio-tagging-2019) in the repository. I tried to clarify some parts of the description, added instructions to reproduce my results, acknowledgments where due, and of course useless badges... 😄 \nFeedback are of course welcomed.", "557622": "“Dad, you are the best and you will be at the very top” . THIS is real motivation !!! Congrats :)", "557659": "Thank for noticing it. All credits goes to him. 😉", "563922": "Note: Updated title and happy to become master 🥳", "573149": "thank you so much for sharing", "573861": "You’re welcomed." }, "source": "meta" }