Huge thanks to Kaggle and DCASE 2018 organizer. It was a great challenge and I really enjoyed participating. Also thanks to other participants, especially to Zafar and daisukelab, who shared great kernels and discussions.
\n\nDuring this challenge, our main aim was to optimize the model efficiently with strong augmentation/regularization (mixup). Our model is based on densenet architecture and several techniques were tried for fast optimization, model ensembles and label noise. I attached our technical report below for detailed explanation.
\n\nCurrently we are preparing for DCASE 2018 workshop paper. Hope to meet you all there.
\n\np.s. dear organizers: I just found in rules page that:
\n\n'Kaggle will notify the potential winner(s) by email. If a potential winner does not respond to the notification attempt within five (5) days from the first notification attempt, then such potential winner will be disqualified...'
\n\n... and I also noticed that my email storage was full. Can you send it again if you already sent it?
", "messages": [ { "id": "365414", "postDate": "08/02/2018 14:19:08", "content": "Huge thanks to Kaggle and DCASE 2018 organizer. It was a great challenge and I really enjoyed participating. Also thanks to other participants, especially to Zafar and daisukelab, who shared great kernels and discussions.
\n\nDuring this challenge, our main aim was to optimize the model efficiently with strong augmentation/regularization (mixup). Our model is based on densenet architecture and several techniques were tried for fast optimization, model ensembles and label noise. I attached our technical report below for detailed explanation.
\n\nCurrently we are preparing for DCASE 2018 workshop paper. Hope to meet you all there.
\n\np.s. dear organizers: I just found in rules page that:
\n\n'Kaggle will notify the potential winner(s) by email. If a potential winner does not respond to the notification attempt within five (5) days from the first notification attempt, then such potential winner will be disqualified...'
\n\n... and I also noticed that my email storage was full. Can you send it again if you already sent it?
", "rawMarkdown": "Huge thanks to Kaggle and DCASE 2018 organizer. It was a great challenge and I really enjoyed participating. Also thanks to other participants, especially to Zafar and daisukelab, who shared great kernels and discussions.\n\nDuring this challenge, our main aim was to optimize the model efficiently with strong augmentation/regularization (mixup). Our model is based on densenet architecture and several techniques were tried for fast optimization, model ensembles and label noise. I attached our technical report below for detailed explanation.\n\nCurrently we are preparing for DCASE 2018 workshop paper. Hope to meet you all there.\n\n\n\np.s. dear organizers: I just found in rules page that:\n\n'Kaggle will notify the potential winner(s) by email. If a potential winner does not respond to the notification attempt within five (5) days from the first notification attempt, then such potential winner will be disqualified...'\n\n... and I also noticed that my email storage was full. Can you send it again if you already sent it?", "votes": null }, { "id": "365488", "postDate": "08/02/2018 19:09:10", "content": "Congratulations on winning the competition! I just finished reading the report. Your classifier module and loss masking technique are quite interesting. I wonder if the latter can be tweaked further and if some theoretical justifications can be made. Looking forward to reading the workshop paper.
\n\nI'd also like to thank the other participants for their contributions; particularly daisukelab for detailing his efforts early on as a top scorer.
", "rawMarkdown": "Congratulations on winning the competition! I just finished reading the report. Your classifier module and loss masking technique are quite interesting. I wonder if the latter can be tweaked further and if some theoretical justifications can be made. Looking forward to reading the workshop paper.\n\nI'd also like to thank the other participants for their contributions; particularly daisukelab for detailing his efforts early on as a top scorer.", "votes": null }, { "id": "365568", "postDate": "08/02/2018 23:05:47", "content": "Dear winner, congrats to you and the Cochlear.ai team on behalf of the Task organizers! We are all eager to read the Workshop paper.
\n\nWe'll look into the issue that you mention about the notification mail!
", "rawMarkdown": "Dear winner, congrats to you and the Cochlear.ai team on behalf of the Task organizers! We are all eager to read the Workshop paper.\n\nWe'll look into the issue that you mention about the notification mail!", "votes": null }, { "id": "365632", "postDate": "08/03/2018 04:35:48", "content": "Hi Il-Young,
\n\nCongratulations! As there is no monetary prize in this competition, we will not be sending out official winners' emails as you've noted. Thanks for participating and great work on your first place finish!
\n\nBest,\nAddison
", "rawMarkdown": "Hi Il-Young,\n\nCongratulations! As there is no monetary prize in this competition, we will not be sending out official winners' emails as you've noted. Thanks for participating and great work on your first place finish!\n\nBest,\nAddison", "votes": null }, { "id": "365699", "postDate": "08/03/2018 07:59:57", "content": "Great job! Congratulations!
", "rawMarkdown": "Great job! Congratulations!", "votes": null }, { "id": "367348", "postDate": "08/07/2018 15:08:30", "content": "Congratulations for winning competition! And thank you for sharing technical report here, I’ve enjoyed reading your solution.
\n\nI've impressed that your system is basically highly automated, class balancing is done by picking one from all classes to form 41 batch-size, in-built log-mel layer to convert raw wave into features online, adaption to mixup and etc.\nAnd the loss masking technique is cool, outliers were always annoyed us so much and your system has auto detection for them.
\n\nI'm looking forward to reading paper, thanks again!
", "rawMarkdown": "Congratulations for winning competition! And thank you for sharing technical report here, I’ve enjoyed reading your solution.\n\nI've impressed that your system is basically highly automated, class balancing is done by picking one from all classes to form 41 batch-size, in-built log-mel layer to convert raw wave into features online, adaption to mixup and etc.\nAnd the loss masking technique is cool, outliers were always annoyed us so much and your system has auto detection for them.\n\nI'm looking forward to reading paper, thanks again!", "votes": null }, { "id": "406284", "postDate": "10/19/2018 01:45:41", "content": "Hello! I am interested in your multi-head softmax system, which you say accelerates the training of your model. Are you able to be more specific about how much improvement in training speed it provides? Did it yield dramatically faster convergence, or only a slight improvement?
", "rawMarkdown": "Hello! I am interested in your multi-head softmax system, which you say accelerates the training of your model. Are you able to be more specific about how much improvement in training speed it provides? Did it yield dramatically faster convergence, or only a slight improvement?", "votes": null } ], "comments": [ { "id": 365488, "author_name": "tqbl95", "author_url": "", "post_date": "08/02/2018 19:09:10", "content": "Congratulations on winning the competition! I just finished reading the report. Your classifier module and loss masking technique are quite interesting. I wonder if the latter can be tweaked further and if some theoretical justifications can be made. Looking forward to reading the workshop paper.
\n\nI'd also like to thank the other participants for their contributions; particularly daisukelab for detailing his efforts early on as a top scorer.
", "votes": null, "replies": [] }, { "id": 365568, "author_name": "eduardofonseca", "author_url": "", "post_date": "08/02/2018 23:05:47", "content": "Dear winner, congrats to you and the Cochlear.ai team on behalf of the Task organizers! We are all eager to read the Workshop paper.
\n\nWe'll look into the issue that you mention about the notification mail!
", "votes": null, "replies": [] }, { "id": 365632, "author_name": "addisonhoward", "author_url": "", "post_date": "08/03/2018 04:35:48", "content": "Hi Il-Young,
\n\nCongratulations! As there is no monetary prize in this competition, we will not be sending out official winners' emails as you've noted. Thanks for participating and great work on your first place finish!
\n\nBest,\nAddison
", "votes": null, "replies": [] }, { "id": 365699, "author_name": "jasonzhang156", "author_url": "", "post_date": "08/03/2018 07:59:57", "content": "Great job! Congratulations!
", "votes": null, "replies": [] }, { "id": 367348, "author_name": "daisukelab", "author_url": "", "post_date": "08/07/2018 15:08:30", "content": "Congratulations for winning competition! And thank you for sharing technical report here, I’ve enjoyed reading your solution.
\n\nI've impressed that your system is basically highly automated, class balancing is done by picking one from all classes to form 41 batch-size, in-built log-mel layer to convert raw wave into features online, adaption to mixup and etc.\nAnd the loss masking technique is cool, outliers were always annoyed us so much and your system has auto detection for them.
\n\nI'm looking forward to reading paper, thanks again!
", "votes": null, "replies": [] }, { "id": 406284, "author_name": "paiforsyth", "author_url": "", "post_date": "10/19/2018 01:45:41", "content": "Hello! I am interested in your multi-head softmax system, which you say accelerates the training of your model. Are you able to be more specific about how much improvement in training speed it provides? Did it yield dramatically faster convergence, or only a slight improvement?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "365414": "Huge thanks to Kaggle and DCASE 2018 organizer. It was a great challenge and I really enjoyed participating. Also thanks to other participants, especially to Zafar and daisukelab, who shared great kernels and discussions.\n\nDuring this challenge, our main aim was to optimize the model efficiently with strong augmentation/regularization (mixup). Our model is based on densenet architecture and several techniques were tried for fast optimization, model ensembles and label noise. I attached our technical report below for detailed explanation.\n\nCurrently we are preparing for DCASE 2018 workshop paper. Hope to meet you all there.\n\n\n\np.s. dear organizers: I just found in rules page that:\n\n'Kaggle will notify the potential winner(s) by email. If a potential winner does not respond to the notification attempt within five (5) days from the first notification attempt, then such potential winner will be disqualified...'\n\n... and I also noticed that my email storage was full. Can you send it again if you already sent it?", "365488": "Congratulations on winning the competition! I just finished reading the report. Your classifier module and loss masking technique are quite interesting. I wonder if the latter can be tweaked further and if some theoretical justifications can be made. Looking forward to reading the workshop paper.\n\nI'd also like to thank the other participants for their contributions; particularly daisukelab for detailing his efforts early on as a top scorer.", "365568": "Dear winner, congrats to you and the Cochlear.ai team on behalf of the Task organizers! We are all eager to read the Workshop paper.\n\nWe'll look into the issue that you mention about the notification mail!", "365632": "Hi Il-Young,\n\nCongratulations! As there is no monetary prize in this competition, we will not be sending out official winners' emails as you've noted. Thanks for participating and great work on your first place finish!\n\nBest,\nAddison", "365699": "Great job! Congratulations!", "367348": "Congratulations for winning competition! And thank you for sharing technical report here, I’ve enjoyed reading your solution.\n\nI've impressed that your system is basically highly automated, class balancing is done by picking one from all classes to form 41 batch-size, in-built log-mel layer to convert raw wave into features online, adaption to mixup and etc.\nAnd the loss masking technique is cool, outliers were always annoyed us so much and your system has auto detection for them.\n\nI'm looking forward to reading paper, thanks again!", "406284": "Hello! I am interested in your multi-head softmax system, which you say accelerates the training of your model. Are you able to be more specific about how much improvement in training speed it provides? Did it yield dramatically faster convergence, or only a slight improvement?" }, "source": "meta" }