I made available my inference kernel.I added comments to make things readable. Hopefully it is useful for fastai users as I saw no inference kernel using fastai library publicly available in this competition.
\n\nI also released the locally trained weights and uploaded them as a public dataset.\nIt gave 12th position on the public leaderboard with a lwlrap .738. It was 8th on the withdrawn private leaderboard (no LB fall please 🙏 😄) .
\n\nMain points:\n- for each audio clip, average predictions of a window sliding on the whole audio clip\n- use 2 different models (3 warm-up, 10 folds)\n- generate the test list as an in-memory CSV (filename+window position), which is much faster than adding rows to a pandas dataframe\n- avoid PIL when loading images
", "messages": [ { "id": "556724", "postDate": "06/20/2019 15:15:55", "content": "I made available my inference kernel.I added comments to make things readable. Hopefully it is useful for fastai users as I saw no inference kernel using fastai library publicly available in this competition.
\n\nI also released the locally trained weights and uploaded them as a public dataset.\nIt gave 12th position on the public leaderboard with a lwlrap .738. It was 8th on the withdrawn private leaderboard (no LB fall please 🙏 😄) .
\n\nMain points:\n- for each audio clip, average predictions of a window sliding on the whole audio clip\n- use 2 different models (3 warm-up, 10 folds)\n- generate the test list as an in-memory CSV (filename+window position), which is much faster than adding rows to a pandas dataframe\n- avoid PIL when loading images
", "rawMarkdown": "I made available my [inference kernel](https://www.kaggle.com/ebouteillon/12th-public-lb-inference-kernel-using-fastai).I added comments to make things readable. Hopefully it is useful for **fastai** users as I saw no inference kernel using fastai library publicly available in this competition.\n\nI also released the locally trained weights and uploaded them as [a public dataset](https://www.kaggle.com/ebouteillon/freesoundaudiotagging2019ebouteillonsolution).\nIt gave 12th position on the public leaderboard with a lwlrap .738. It was 8th on the withdrawn private leaderboard (no LB fall please 🙏 😄) .\n\n**Main points:**\n- for each audio clip, average predictions of a window sliding on the whole audio clip\n- use 2 different models (3 warm-up, 10 folds)\n- generate the test list as an in-memory CSV (filename+window position), which is much faster than adding rows to a pandas dataframe\n- avoid PIL when loading images", "votes": null }, { "id": "556735", "postDate": "06/20/2019 15:26:08", "content": "We use fastai approach too :D, but it is not good enough. \nBTW, congrats your private position. I believe that it will not be changed
Thanks for sharing the solution...I also used fastai, but not as good as yours. Also when the withdrawn private LB was announced it was past midnight for me, so never got to saw my initial position :(
\n\nI don't think you would see too much of a change in your private LB position...good luck!!
", "rawMarkdown": "Thanks for sharing the solution...I also used fastai, but not as good as yours. Also when the withdrawn private LB was announced it was past midnight for me, so never got to saw my initial position :( \n\nI don't think you would see too much of a change in your private LB position...good luck!!", "votes": null }, { "id": "563930", "postDate": "06/28/2019 19:33:32", "content": "Note: updated title and happy to get master. 🥳
", "rawMarkdown": "Note: updated title and happy to get master. 🥳", "votes": null }, { "id": "564231", "postDate": "06/29/2019 06:53:16", "content": "Congratulations @ebouteillon ! Next stop: GM :-)\nAnd thanks for sharing your solution.
", "rawMarkdown": "Congratulations @ebouteillon ! Next stop: GM :-)\nAnd thanks for sharing your solution.", "votes": null } ], "comments": [ { "id": 556735, "author_name": "backaggle", "author_url": "", "post_date": "06/20/2019 15:26:08", "content": "We use fastai approach too :D, but it is not good enough. \nBTW, congrats your private position. I believe that it will not be changed
Thanks for sharing the solution...I also used fastai, but not as good as yours. Also when the withdrawn private LB was announced it was past midnight for me, so never got to saw my initial position :(
\n\nI don't think you would see too much of a change in your private LB position...good luck!!
", "votes": null, "replies": [] }, { "id": 563930, "author_name": "ebouteillon", "author_url": "", "post_date": "06/28/2019 19:33:32", "content": "Note: updated title and happy to get master. 🥳
", "votes": null, "replies": [] }, { "id": 564231, "author_name": "rohanrao", "author_url": "", "post_date": "06/29/2019 06:53:16", "content": "Congratulations @ebouteillon ! Next stop: GM :-)\nAnd thanks for sharing your solution.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "556724": "I made available my [inference kernel](https://www.kaggle.com/ebouteillon/12th-public-lb-inference-kernel-using-fastai).I added comments to make things readable. Hopefully it is useful for **fastai** users as I saw no inference kernel using fastai library publicly available in this competition.\n\nI also released the locally trained weights and uploaded them as [a public dataset](https://www.kaggle.com/ebouteillon/freesoundaudiotagging2019ebouteillonsolution).\nIt gave 12th position on the public leaderboard with a lwlrap .738. It was 8th on the withdrawn private leaderboard (no LB fall please 🙏 😄) .\n\n**Main points:**\n- for each audio clip, average predictions of a window sliding on the whole audio clip\n- use 2 different models (3 warm-up, 10 folds)\n- generate the test list as an in-memory CSV (filename+window position), which is much faster than adding rows to a pandas dataframe\n- avoid PIL when loading images", "556735": "We use `fastai` approach too :D, but it is not good enough. \nBTW, congrats your private position. I believe that it will not be changed", "556813": "Thanks for sharing the solution...I also used fastai, but not as good as yours. Also when the withdrawn private LB was announced it was past midnight for me, so never got to saw my initial position :( \n\nI don't think you would see too much of a change in your private LB position...good luck!!", "563930": "Note: updated title and happy to get master. 🥳", "564231": "Congratulations @ebouteillon ! Next stop: GM :-)\nAnd thanks for sharing your solution." }, "source": "meta" }