As far as I can remember correctly, CNN dominated the last competition. Here's what I've done so far:
\n\nHere's the links.
\n\nAs far as I can remember correctly, CNN dominated the last competition. Here's what I've done so far:
\n\nHere's the links.
\n\nAre you using fast.ai's pre-trained classifiers? Just a reminder that the rules of this challenge prohibit the use of external data or pre-trained models. https://www.kaggle.com/c/freesound-audio-tagging-2019/overview/kernels-requirements
", "rawMarkdown": "Are you using fast.ai's pre-trained classifiers? Just a reminder that the rules of this challenge prohibit the use of external data or pre-trained models. https://www.kaggle.com/c/freesound-audio-tagging-2019/overview/kernels-requirements", "votes": null }, { "id": "508407", "postDate": "04/06/2019 05:59:46", "content": "No pre-trained model, from scratch in less than 30 min.
\n\nI'm following the rule strictly, and that's what I'm asking here:\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064#508071
\n\nOh excuse me, you are the competition host. Then the answer to my question would be clear, pre-trained models are NOT allowed strictly even when loading from datasets.
\n\nI hope this Kernels Requirements would be clear for that:
\n\nI appreciate if you could confirm my understanding is correct. Thank you.
", "rawMarkdown": "No pre-trained model, from scratch in less than 30 min.\n\nI'm following the rule strictly, and that's what I'm asking here:\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064#508071\n\nOh excuse me, you are the competition host. Then the answer to my question would be clear, pre-trained models are NOT allowed strictly even when loading from datasets.\n\nI hope this [Kernels Requirements](https://www.kaggle.com/c/freesound-audio-tagging-2019/overview/kernels-requirements) would be clear for that:\n\n- \"External data and pre-trained models are not allowed. If you train your model offline, you may upload your model as a Kaggle dataset for use in creating your inference Kernel.\" --- This would mean that __you can use your model trained from scratch offline and put it available as (maybe public) dataset.__\n\nI appreciate if you could confirm my understanding is correct. Thank you.", "votes": null }, { "id": "508430", "postDate": "04/06/2019 06:42:39", "content": "Hi, daisukelab! Glad to see you again here!
", "rawMarkdown": "Hi, daisukelab! Glad to see you again here!", "votes": null }, { "id": "508435", "postDate": "04/06/2019 06:46:22", "content": "Hi, great to see you too! I'm very curious how you solve multi-label problem. Let's enjoy :)
", "rawMarkdown": "Hi, great to see you too! I'm very curious how you solve multi-label problem. Let's enjoy :)", "votes": null }, { "id": "508437", "postDate": "04/06/2019 06:47:42", "content": "The rules about pretrained model and offline training sound weird and confusing to me. What does \"pretrained\" mean? Does it mean 1) models trained on other datasets, e.g. ImageNet or 2) models trained by others or both 1) and 2) - or once I stopped training, the resulting model would be considered pretrained and should not be used for further training? A more detailed explanation including examples of what is allowed and what is not would be appreciated.
\n\nAs mentioned in the rules discussion thread, there seems no way to enforce the \"pretrain\" rule on kernels that use an offline trained model and do no win the competition. Should this also be addressed?
", "rawMarkdown": "The rules about pretrained model and offline training sound weird and confusing to me. What does \"pretrained\" mean? Does it mean 1) models trained on other datasets, e.g. ImageNet or 2) models trained by others or both 1) and 2) - or once I stopped training, the resulting model would be considered pretrained and should not be used for further training? A more detailed explanation including examples of what is allowed and what is not would be appreciated.\n\nAs mentioned in the rules discussion thread, there seems no way to enforce the \"pretrain\" rule on kernels that use an offline trained model and do no win the competition. Should this also be addressed?", "votes": null }, { "id": "509559", "postDate": "04/08/2019 00:54:31", "content": "Why do you use 3 dimensional images since pretrain network are not allowed?\nIt seems that the 3 dimensions are only a stack of the same dimension.\nI wonder if this problem is a multi-label problem since the perfect score is when you put only one 1 in the good class.
", "rawMarkdown": "Why do you use 3 dimensional images since pretrain network are not allowed?\nIt seems that the 3 dimensions are only a stack of the same dimension.\nI wonder if this problem is a multi-label problem since the perfect score is when you put only one 1 in the good class.", "votes": null }, { "id": "509564", "postDate": "04/08/2019 01:16:42", "content": "Two reasons:\n- No need to modify tested/verified genuine models. Easy to switch between models is also beneficial.\n- Multi-channel leaves space for adding/exchanging data for example; ch.1 for spectrogram, ch.2 for temporal difference, ch.3 for ...
", "rawMarkdown": "Two reasons:\n- No need to modify tested/verified genuine models. Easy to switch between models is also beneficial.\n- Multi-channel leaves space for adding/exchanging data for example; ch.1 for spectrogram, ch.2 for temporal difference, ch.3 for ...", "votes": null }, { "id": "510212", "postDate": "04/08/2019 20:26:23", "content": "Just to close out the discussion, please refer to this thread for a discussion of the submission rules\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064
\n\nre enforcement on what is used for offline training: as noted elsewhere, we are relying on a combination of the honor system as well as a manual review of the code for the top ranked submissions. This worked well for us last year where had a purely offline (no kernels) challenge and we allowed external data and pre-trained models but with restrictions that we enforced at review time.
", "rawMarkdown": "Just to close out the discussion, please refer to this thread for a discussion of the submission rules\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064\n\nre enforcement on what is used for offline training: as noted elsewhere, we are relying on a combination of the honor system as well as a manual review of the code for the top ranked submissions. This worked well for us last year where had a purely offline (no kernels) challenge and we allowed external data and pre-trained models but with restrictions that we enforced at review time.", "votes": null }, { "id": "510262", "postDate": "04/08/2019 22:18:09", "content": "Dear Competition Host, I would like to know what are exactly the limitations for the training in terms of computation time and memory. Is it 4hours with CPU starting from scratch? Which means that it that case the choice of the library is very important. For instance, Pytorch will be much better than Keras since it more less 2x time efficient.
", "rawMarkdown": "Dear Competition Host, I would like to know what are exactly the limitations for the training in terms of computation time and memory. Is it 4hours with CPU starting from scratch? Which means that it that case the choice of the library is very important. For instance, Pytorch will be much better than Keras since it more less 2x time efficient.", "votes": null }, { "id": "510278", "postDate": "04/08/2019 22:55:03", "content": "@pierretisseur do you mind asking this question again in the official thread so that it will benefit more people who read that thread? It's better to have all rules discussion in that thread. Thanks.\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064
", "rawMarkdown": "pierretisseur do you mind asking this question again in the official thread so that it will benefit more people who read that thread? It's better to have all rules discussion in that thread. Thanks.\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064", "votes": null } ], "comments": [ { "id": 508405, "author_name": "plakal", "author_url": "", "post_date": "04/06/2019 05:47:42", "content": "Are you using fast.ai's pre-trained classifiers? Just a reminder that the rules of this challenge prohibit the use of external data or pre-trained models. https://www.kaggle.com/c/freesound-audio-tagging-2019/overview/kernels-requirements
", "votes": null, "replies": [ { "id": 508407, "author_name": "daisukelab", "author_url": "", "post_date": "04/06/2019 05:59:46", "content": "No pre-trained model, from scratch in less than 30 min.
\n\nI'm following the rule strictly, and that's what I'm asking here:\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064#508071
\n\nOh excuse me, you are the competition host. Then the answer to my question would be clear, pre-trained models are NOT allowed strictly even when loading from datasets.
\n\nI hope this Kernels Requirements would be clear for that:
\n\nI appreciate if you could confirm my understanding is correct. Thank you.
", "votes": null, "replies": [] }, { "id": 508437, "author_name": "ritsu1228", "author_url": "", "post_date": "04/06/2019 06:47:42", "content": "The rules about pretrained model and offline training sound weird and confusing to me. What does \"pretrained\" mean? Does it mean 1) models trained on other datasets, e.g. ImageNet or 2) models trained by others or both 1) and 2) - or once I stopped training, the resulting model would be considered pretrained and should not be used for further training? A more detailed explanation including examples of what is allowed and what is not would be appreciated.
\n\nAs mentioned in the rules discussion thread, there seems no way to enforce the \"pretrain\" rule on kernels that use an offline trained model and do no win the competition. Should this also be addressed?
", "votes": null, "replies": [] }, { "id": 510212, "author_name": "plakal", "author_url": "", "post_date": "04/08/2019 20:26:23", "content": "Just to close out the discussion, please refer to this thread for a discussion of the submission rules\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064
\n\nre enforcement on what is used for offline training: as noted elsewhere, we are relying on a combination of the honor system as well as a manual review of the code for the top ranked submissions. This worked well for us last year where had a purely offline (no kernels) challenge and we allowed external data and pre-trained models but with restrictions that we enforced at review time.
", "votes": null, "replies": [] }, { "id": 510262, "author_name": "pierretisseur", "author_url": "", "post_date": "04/08/2019 22:18:09", "content": "Dear Competition Host, I would like to know what are exactly the limitations for the training in terms of computation time and memory. Is it 4hours with CPU starting from scratch? Which means that it that case the choice of the library is very important. For instance, Pytorch will be much better than Keras since it more less 2x time efficient.
", "votes": null, "replies": [] }, { "id": 510278, "author_name": "plakal", "author_url": "", "post_date": "04/08/2019 22:55:03", "content": "@pierretisseur do you mind asking this question again in the official thread so that it will benefit more people who read that thread? It's better to have all rules discussion in that thread. Thanks.\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064
", "votes": null, "replies": [] } ] }, { "id": 508430, "author_name": "kelexu", "author_url": "", "post_date": "04/06/2019 06:42:39", "content": "Hi, daisukelab! Glad to see you again here!
", "votes": null, "replies": [ { "id": 508435, "author_name": "daisukelab", "author_url": "", "post_date": "04/06/2019 06:46:22", "content": "Hi, great to see you too! I'm very curious how you solve multi-label problem. Let's enjoy :)
", "votes": null, "replies": [] } ] }, { "id": 509559, "author_name": "pierretisseur", "author_url": "", "post_date": "04/08/2019 00:54:31", "content": "Why do you use 3 dimensional images since pretrain network are not allowed?\nIt seems that the 3 dimensions are only a stack of the same dimension.\nI wonder if this problem is a multi-label problem since the perfect score is when you put only one 1 in the good class.
", "votes": null, "replies": [ { "id": 509564, "author_name": "daisukelab", "author_url": "", "post_date": "04/08/2019 01:16:42", "content": "Two reasons:\n- No need to modify tested/verified genuine models. Easy to switch between models is also beneficial.\n- Multi-channel leaves space for adding/exchanging data for example; ch.1 for spectrogram, ch.2 for temporal difference, ch.3 for ...
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "508391": "As far as I can remember correctly, CNN dominated the last competition. Here's what I've done so far:\n\n- Convert all curated samples into 2D images on memory. [128, 128*audio-length, 3]\n- Just followed fast.ai multi-label image classifier solution, yes it is just same as image classification.\n\nHere's the links.\n\n- [Kernel for this competition](https://www.kaggle.com/daisukelab/cnn-2d-basic-solution-powered-by-fast-ai)\n Current score is still simple try, much space left to improve.\n- [My solution for the last 2018 competition](https://www.kaggle.com/daisukelab/freesound-dataset-kaggle-2018-solution)\n Useless for this competition due to training time restriction.\n\n## Update on 30-Apr, 2019\n- lwlrap wrapper for fast.ai is ready.\n- TTA is applied --> but maybe it is not working fine for this use case...", "508405": "Are you using fast.ai's pre-trained classifiers? Just a reminder that the rules of this challenge prohibit the use of external data or pre-trained models. https://www.kaggle.com/c/freesound-audio-tagging-2019/overview/kernels-requirements", "508407": "No pre-trained model, from scratch in less than 30 min.\n\nI'm following the rule strictly, and that's what I'm asking here:\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064#508071\n\nOh excuse me, you are the competition host. Then the answer to my question would be clear, pre-trained models are NOT allowed strictly even when loading from datasets.\n\nI hope this [Kernels Requirements](https://www.kaggle.com/c/freesound-audio-tagging-2019/overview/kernels-requirements) would be clear for that:\n\n- \"External data and pre-trained models are not allowed. If you train your model offline, you may upload your model as a Kaggle dataset for use in creating your inference Kernel.\" --- This would mean that __you can use your model trained from scratch offline and put it available as (maybe public) dataset.__\n\nI appreciate if you could confirm my understanding is correct. Thank you.", "508430": "Hi, daisukelab! Glad to see you again here!", "508435": "Hi, great to see you too! I'm very curious how you solve multi-label problem. Let's enjoy :)", "508437": "The rules about pretrained model and offline training sound weird and confusing to me. What does \"pretrained\" mean? Does it mean 1) models trained on other datasets, e.g. ImageNet or 2) models trained by others or both 1) and 2) - or once I stopped training, the resulting model would be considered pretrained and should not be used for further training? A more detailed explanation including examples of what is allowed and what is not would be appreciated.\n\nAs mentioned in the rules discussion thread, there seems no way to enforce the \"pretrain\" rule on kernels that use an offline trained model and do no win the competition. Should this also be addressed?", "509559": "Why do you use 3 dimensional images since pretrain network are not allowed?\nIt seems that the 3 dimensions are only a stack of the same dimension.\nI wonder if this problem is a multi-label problem since the perfect score is when you put only one 1 in the good class.", "509564": "Two reasons:\n- No need to modify tested/verified genuine models. Easy to switch between models is also beneficial.\n- Multi-channel leaves space for adding/exchanging data for example; ch.1 for spectrogram, ch.2 for temporal difference, ch.3 for ...", "510212": "Just to close out the discussion, please refer to this thread for a discussion of the submission rules\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064\n\nre enforcement on what is used for offline training: as noted elsewhere, we are relying on a combination of the honor system as well as a manual review of the code for the top ranked submissions. This worked well for us last year where had a purely offline (no kernels) challenge and we allowed external data and pre-trained models but with restrictions that we enforced at review time.", "510262": "Dear Competition Host, I would like to know what are exactly the limitations for the training in terms of computation time and memory. Is it 4hours with CPU starting from scratch? Which means that it that case the choice of the library is very important. For instance, Pytorch will be much better than Keras since it more less 2x time efficient.", "510278": "pierretisseur do you mind asking this question again in the official thread so that it will benefit more people who read that thread? It's better to have all rules discussion in that thread. Thanks.\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/88064" }, "source": "meta" }