First of all, congratulations everyone! And thanks to the organizers - this was a really well organized competition. Even though I joined quite late (I think around 3 weeks back) and this was my first venture into the Audio domain, my learning curve has been incredibly high! I look forward to continue my work in this domain.
\n\nMy best single model with nearly zero pre-processing scores 0.915 in the private LB.
\n\nAs part of pre-processing, only silence removal and chunking was used. (No label refinement, no mixup).
\n\nPrivate LB: 0.915
\n\nPublic LB: 0.921
\n\nModel: CNN
\n\nNumber of Parameters: Less than 1.5M
\n\nNumber of folds: 5
\n\nI think apart from the architecture, its strength comes from the feature engineering, which followed a little-bit different approach than what has been discussed here in the forums.
\n\nThe same process with mixup scored 0.907 in Private LB, 0.925 in Public LB.
\n\nI wish I had submitted these results in the technical report and final submission :'(
", "messages": [ { "id": "364884", "postDate": "08/01/2018 12:58:13", "content": "First of all, congratulations everyone! And thanks to the organizers - this was a really well organized competition. Even though I joined quite late (I think around 3 weeks back) and this was my first venture into the Audio domain, my learning curve has been incredibly high! I look forward to continue my work in this domain.
\n\nMy best single model with nearly zero pre-processing scores 0.915 in the private LB.
\n\nAs part of pre-processing, only silence removal and chunking was used. (No label refinement, no mixup).
\n\nPrivate LB: 0.915
\n\nPublic LB: 0.921
\n\nModel: CNN
\n\nNumber of Parameters: Less than 1.5M
\n\nNumber of folds: 5
\n\nI think apart from the architecture, its strength comes from the feature engineering, which followed a little-bit different approach than what has been discussed here in the forums.
\n\nThe same process with mixup scored 0.907 in Private LB, 0.925 in Public LB.
\n\nI wish I had submitted these results in the technical report and final submission :'(
", "rawMarkdown": "First of all, congratulations everyone! And thanks to the organizers - this was a really well organized competition. Even though I joined quite late (I think around 3 weeks back) and this was my first venture into the Audio domain, my learning curve has been incredibly high! I look forward to continue my work in this domain. \n\nMy best single model with nearly zero pre-processing scores 0.915 in the private LB. \n\nAs part of pre-processing, only silence removal and chunking was used. (No label refinement, no mixup).\n\nPrivate LB: 0.915\n\nPublic LB: 0.921\n\nModel: CNN\n\nNumber of Parameters: Less than 1.5M\n\nNumber of folds: 5\n\nI think apart from the architecture, its strength comes from the feature engineering, which followed a little-bit different approach than what has been discussed here in the forums. \n\nThe same process with mixup scored 0.907 in Private LB, 0.925 in Public LB. \n\nI wish I had submitted these results in the technical report and final submission :'(", "votes": null }, { "id": "364889", "postDate": "08/01/2018 13:16:56", "content": "Good job, with only 3 weeks!\nSeems public and private rank are so different for some guys
", "rawMarkdown": "Good job, with only 3 weeks!\nSeems public and private rank are so different for some guys", "votes": null }, { "id": "364896", "postDate": "08/01/2018 13:33:25", "content": "Yes, there has been some pretty evident LB shake-up. Although I did anticipate this - given how the Public and Private LB split has been performed...
", "rawMarkdown": "Yes, there has been some pretty evident LB shake-up. Although I did anticipate this - given how the Public and Private LB split has been performed...", "votes": null }, { "id": "365052", "postDate": "08/01/2018 19:48:21", "content": "My best single model is a double GRU on MFCC. 200k parameters. Private LB score 0.869, public LB 0.906. https://github.com/artyompal/kaggle_freesound/blob/master/v85_gru.py
", "rawMarkdown": "My best single model is a double GRU on MFCC. 200k parameters. Private LB score 0.869, public LB 0.906. https://github.com/artyompal/kaggle_freesound/blob/master/v85_gru.py", "votes": null }, { "id": "367214", "postDate": "08/07/2018 10:28:27", "content": "My best single model:\nPrivate LB: 0.9413\nPublic LB: 0.9251\nModel CNN (VGG-like with 2 inputs: 1. log-mel, 2.mean and standard deviation of log-mel)\nData augmentation: pitch shift, time stretch, mix-up, erase\nRegularization: label smoothing\nNumber of parameters: around 700K
", "rawMarkdown": "My best single model:\nPrivate LB: 0.9413\nPublic LB: 0.9251\nModel CNN (VGG-like with 2 inputs: 1. log-mel, 2.mean and standard deviation of log-mel)\nData augmentation: pitch shift, time stretch, mix-up, erase\nRegularization: label smoothing\nNumber of parameters: around 700K", "votes": null } ], "comments": [ { "id": 364889, "author_name": "sailorwei", "author_url": "", "post_date": "08/01/2018 13:16:56", "content": "Good job, with only 3 weeks!\nSeems public and private rank are so different for some guys
", "votes": null, "replies": [ { "id": 364896, "author_name": "gyat2017", "author_url": "", "post_date": "08/01/2018 13:33:25", "content": "Yes, there has been some pretty evident LB shake-up. Although I did anticipate this - given how the Public and Private LB split has been performed...
", "votes": null, "replies": [] } ] }, { "id": 365052, "author_name": "artyomp", "author_url": "", "post_date": "08/01/2018 19:48:21", "content": "My best single model is a double GRU on MFCC. 200k parameters. Private LB score 0.869, public LB 0.906. https://github.com/artyompal/kaggle_freesound/blob/master/v85_gru.py
", "votes": null, "replies": [] }, { "id": 367214, "author_name": "thomeou", "author_url": "", "post_date": "08/07/2018 10:28:27", "content": "My best single model:\nPrivate LB: 0.9413\nPublic LB: 0.9251\nModel CNN (VGG-like with 2 inputs: 1. log-mel, 2.mean and standard deviation of log-mel)\nData augmentation: pitch shift, time stretch, mix-up, erase\nRegularization: label smoothing\nNumber of parameters: around 700K
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "364884": "First of all, congratulations everyone! And thanks to the organizers - this was a really well organized competition. Even though I joined quite late (I think around 3 weeks back) and this was my first venture into the Audio domain, my learning curve has been incredibly high! I look forward to continue my work in this domain. \n\nMy best single model with nearly zero pre-processing scores 0.915 in the private LB. \n\nAs part of pre-processing, only silence removal and chunking was used. (No label refinement, no mixup).\n\nPrivate LB: 0.915\n\nPublic LB: 0.921\n\nModel: CNN\n\nNumber of Parameters: Less than 1.5M\n\nNumber of folds: 5\n\nI think apart from the architecture, its strength comes from the feature engineering, which followed a little-bit different approach than what has been discussed here in the forums. \n\nThe same process with mixup scored 0.907 in Private LB, 0.925 in Public LB. \n\nI wish I had submitted these results in the technical report and final submission :'(", "364889": "Good job, with only 3 weeks!\nSeems public and private rank are so different for some guys", "364896": "Yes, there has been some pretty evident LB shake-up. Although I did anticipate this - given how the Public and Private LB split has been performed...", "365052": "My best single model is a double GRU on MFCC. 200k parameters. Private LB score 0.869, public LB 0.906. https://github.com/artyompal/kaggle_freesound/blob/master/v85_gru.py", "367214": "My best single model:\nPrivate LB: 0.9413\nPublic LB: 0.9251\nModel CNN (VGG-like with 2 inputs: 1. log-mel, 2.mean and standard deviation of log-mel)\nData augmentation: pitch shift, time stretch, mix-up, erase\nRegularization: label smoothing\nNumber of parameters: around 700K" }, "source": "meta" }