Capsule Nets gave some good results in the Toxic Comment Classification Challenge.
\n\nI Used this kernel from Toxic Comment Classification Challenge to apply the same model to this challange.
\n\nHave a look: https://www.kaggle.com/ghostiphate/capsule-net !
", "messages": [ { "id": "420949", "postDate": "11/14/2018 11:39:23", "content": "Capsule Nets gave some good results in the Toxic Comment Classification Challenge.
\n\nI Used this kernel from Toxic Comment Classification Challenge to apply the same model to this challange.
\n\nHave a look: https://www.kaggle.com/ghostiphate/capsule-net !
", "rawMarkdown": "Capsule Nets gave some good results in the Toxic Comment Classification Challenge.\n\nI Used [this kernel][1] from Toxic Comment Classification Challenge to apply the same model to this challange. \n\nHave a look: https://www.kaggle.com/ghostiphate/capsule-net !\n\n\n [1]: https://www.kaggle.com/chongjiujjin/capsule-net-with-gru", "votes": null }, { "id": "420971", "postDate": "11/14/2018 12:17:38", "content": "Sorry but your link seems to be broken
", "rawMarkdown": "Sorry but your link seems to be broken", "votes": null }, { "id": "420982", "postDate": "11/14/2018 12:33:07", "content": "Capsules can be computationnally heavy to be really effective (by tuning properly the parameters)
\n\nBeware of time limitation constraint in this competition.
", "rawMarkdown": "Capsules can be computationnally heavy to be really effective (by tuning properly the parameters) \n\nBeware of time limitation constraint in this competition.", "votes": null }, { "id": "420992", "postDate": "11/14/2018 12:41:24", "content": "You need to remove the exclamation mark ! at the end of the link
", "rawMarkdown": "You need to remove the exclamation mark ! at the end of the link", "votes": null }, { "id": "421096", "postDate": "11/14/2018 15:19:58", "content": "fixed
", "rawMarkdown": "fixed", "votes": null }, { "id": "421181", "postDate": "11/14/2018 17:12:20", "content": "Great point. CapNet is slow. I would like to use the time to run several CNN models and blend them.
", "rawMarkdown": "Great point. CapNet is slow. I would like to use the time to run several CNN models and blend them.", "votes": null }, { "id": "421205", "postDate": "11/14/2018 18:04:00", "content": "Interesting !!
\n\nAs for me , I still use single model with k-folds ....It remains about 1200s. May be I need to optimize my model to be able to run and blend others models ( or stick with single fold )
", "rawMarkdown": "Interesting !!\n\nAs for me , I still use single model with k-folds ....It remains about 1200s. May be I need to optimize my model to be able to run and blend others models ( or stick with single fold )", "votes": null }, { "id": "421221", "postDate": "11/14/2018 18:35:30", "content": "Wow, that is a pretty impressive result with a single model.
", "rawMarkdown": "Wow, that is a pretty impressive result with a single model.", "votes": null }, { "id": "421236", "postDate": "11/14/2018 19:13:54", "content": "@serigne, super cool! I plan to refactorize my code to make it more efficient.
", "rawMarkdown": "serigne, super cool! I plan to refactorize my code to make it more efficient.", "votes": null }, { "id": "421262", "postDate": "11/14/2018 20:11:39", "content": "I've also stuck with a single model at this point. Iterating through a bunch of different architectures to find a \"good\" candidate, but given the time constraints and my lack of knowing how to \"optimize\" the kernels, I'll most likely do the same as @serigne.
", "rawMarkdown": "I've also stuck with a single model at this point. Iterating through a bunch of different architectures to find a \"good\" candidate, but given the time constraints and my lack of knowing how to \"optimize\" the kernels, I'll most likely do the same as @serigne.", "votes": null }, { "id": "421275", "postDate": "11/14/2018 20:42:55", "content": "@Sreeram Thanks !\nHowever , we are in early days of 3 months competition. I guess people will have enough time to get insight on the data and build much better models :)
", "rawMarkdown": "Sreeram Thanks !\nHowever , we are in early days of 3 months competition. I guess people will have enough time to get insight on the data and build much better models :)", "votes": null }, { "id": "421276", "postDate": "11/14/2018 20:46:27", "content": "@Learnmower I tried some different architectures too and they stopped improving after 3 or 4 epoches.
\n\nNow it seems I've found a relatively \"good candidate\" (at least which works better than the others I tried)
", "rawMarkdown": "Learnmower I tried some different architectures too and they stopped improving after 3 or 4 epoches. \n\nNow it seems I've found a relatively \"good candidate\" (at least which works better than the others I tried)", "votes": null }, { "id": "421284", "postDate": "11/14/2018 21:16:39", "content": "@Serigne sounds about right on my end, too. 3-4 epochs depending on the architecture. I'm definitely using \"good\" loosely here. In Toxic, I think we traded off stability for score, but it was the blending later that helped to stabilize the variance.
", "rawMarkdown": "Serigne sounds about right on my end, too. 3-4 epochs depending on the architecture. I'm definitely using \"good\" loosely here. In Toxic, I think we traded off stability for score, but it was the blending later that helped to stabilize the variance.", "votes": null }, { "id": "421309", "postDate": "11/14/2018 22:14:55", "content": "\n\n\nI'm definitely using \"good\" loosely here
\n
Same for me....\"good\" is too relative in this phase of the competition :)
", "rawMarkdown": "> I'm definitely using \"good\" loosely here\n\nSame for me....\"good\" is too relative in this phase of the competition :)", "votes": null }, { "id": "421710", "postDate": "11/15/2018 09:57:36", "content": "@serigne, very good score by just one model.\nCan you please tell,\nhow many epochs are you training your single model for each fold?
\n\nAnd 1200s in total or per fold?
\n\nThanks.
", "rawMarkdown": "@serigne, very good score by just one model.\nCan you please tell,\nhow many epochs are you training your single model **for each fold**?\n\nAnd 1200s in total or per fold?\n\nThanks.", "votes": null }, { "id": "422099", "postDate": "11/15/2018 19:39:12", "content": "@prashantkikani about 8 or 9 epoches for each fold
\n\n1200s was remaining time after all training .
\n\nActually I miscalculated. The entire model ( all folds) takes about 5800s. So it remains 1400s
", "rawMarkdown": "prashantkikani about 8 or 9 epoches for each fold \n\n1200s was remaining time after all training . \n\nActually I miscalculated. The entire model ( all folds) takes about 5800s. So it remains 1400s", "votes": null } ], "comments": [ { "id": 420971, "author_name": "demonplus", "author_url": "", "post_date": "11/14/2018 12:17:38", "content": "Sorry but your link seems to be broken
", "votes": null, "replies": [ { "id": 420992, "author_name": "serigne", "author_url": "", "post_date": "11/14/2018 12:41:24", "content": "You need to remove the exclamation mark ! at the end of the link
", "votes": null, "replies": [] }, { "id": 421096, "author_name": "ghostiphate", "author_url": "", "post_date": "11/14/2018 15:19:58", "content": "fixed
", "votes": null, "replies": [] } ] }, { "id": 420982, "author_name": "serigne", "author_url": "", "post_date": "11/14/2018 12:33:07", "content": "Capsules can be computationnally heavy to be really effective (by tuning properly the parameters)
\n\nBeware of time limitation constraint in this competition.
", "votes": null, "replies": [ { "id": 421181, "author_name": "shujian", "author_url": "", "post_date": "11/14/2018 17:12:20", "content": "Great point. CapNet is slow. I would like to use the time to run several CNN models and blend them.
", "votes": null, "replies": [] }, { "id": 421205, "author_name": "serigne", "author_url": "", "post_date": "11/14/2018 18:04:00", "content": "Interesting !!
\n\nAs for me , I still use single model with k-folds ....It remains about 1200s. May be I need to optimize my model to be able to run and blend others models ( or stick with single fold )
", "votes": null, "replies": [] }, { "id": 421221, "author_name": "sreeram1234", "author_url": "", "post_date": "11/14/2018 18:35:30", "content": "Wow, that is a pretty impressive result with a single model.
", "votes": null, "replies": [] }, { "id": 421236, "author_name": "shujian", "author_url": "", "post_date": "11/14/2018 19:13:54", "content": "@serigne, super cool! I plan to refactorize my code to make it more efficient.
", "votes": null, "replies": [] }, { "id": 421262, "author_name": "learnmower", "author_url": "", "post_date": "11/14/2018 20:11:39", "content": "I've also stuck with a single model at this point. Iterating through a bunch of different architectures to find a \"good\" candidate, but given the time constraints and my lack of knowing how to \"optimize\" the kernels, I'll most likely do the same as @serigne.
", "votes": null, "replies": [] }, { "id": 421275, "author_name": "serigne", "author_url": "", "post_date": "11/14/2018 20:42:55", "content": "@Sreeram Thanks !\nHowever , we are in early days of 3 months competition. I guess people will have enough time to get insight on the data and build much better models :)
", "votes": null, "replies": [] }, { "id": 421276, "author_name": "serigne", "author_url": "", "post_date": "11/14/2018 20:46:27", "content": "@Learnmower I tried some different architectures too and they stopped improving after 3 or 4 epoches.
\n\nNow it seems I've found a relatively \"good candidate\" (at least which works better than the others I tried)
", "votes": null, "replies": [] }, { "id": 421284, "author_name": "learnmower", "author_url": "", "post_date": "11/14/2018 21:16:39", "content": "@Serigne sounds about right on my end, too. 3-4 epochs depending on the architecture. I'm definitely using \"good\" loosely here. In Toxic, I think we traded off stability for score, but it was the blending later that helped to stabilize the variance.
", "votes": null, "replies": [] }, { "id": 421309, "author_name": "serigne", "author_url": "", "post_date": "11/14/2018 22:14:55", "content": "\n\n\nI'm definitely using \"good\" loosely here
\n
Same for me....\"good\" is too relative in this phase of the competition :)
", "votes": null, "replies": [] }, { "id": 421710, "author_name": "prashantkikani", "author_url": "", "post_date": "11/15/2018 09:57:36", "content": "@serigne, very good score by just one model.\nCan you please tell,\nhow many epochs are you training your single model for each fold?
\n\nAnd 1200s in total or per fold?
\n\nThanks.
", "votes": null, "replies": [] }, { "id": 422099, "author_name": "serigne", "author_url": "", "post_date": "11/15/2018 19:39:12", "content": "@prashantkikani about 8 or 9 epoches for each fold
\n\n1200s was remaining time after all training .
\n\nActually I miscalculated. The entire model ( all folds) takes about 5800s. So it remains 1400s
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "420949": "Capsule Nets gave some good results in the Toxic Comment Classification Challenge.\n\nI Used [this kernel][1] from Toxic Comment Classification Challenge to apply the same model to this challange. \n\nHave a look: https://www.kaggle.com/ghostiphate/capsule-net !\n\n\n [1]: https://www.kaggle.com/chongjiujjin/capsule-net-with-gru", "420971": "Sorry but your link seems to be broken", "420982": "Capsules can be computationnally heavy to be really effective (by tuning properly the parameters) \n\nBeware of time limitation constraint in this competition.", "420992": "You need to remove the exclamation mark ! at the end of the link", "421096": "fixed", "421181": "Great point. CapNet is slow. I would like to use the time to run several CNN models and blend them.", "421205": "Interesting !!\n\nAs for me , I still use single model with k-folds ....It remains about 1200s. May be I need to optimize my model to be able to run and blend others models ( or stick with single fold )", "421221": "Wow, that is a pretty impressive result with a single model.", "421236": "serigne, super cool! I plan to refactorize my code to make it more efficient.", "421262": "I've also stuck with a single model at this point. Iterating through a bunch of different architectures to find a \"good\" candidate, but given the time constraints and my lack of knowing how to \"optimize\" the kernels, I'll most likely do the same as @serigne.", "421275": "Sreeram Thanks !\nHowever , we are in early days of 3 months competition. I guess people will have enough time to get insight on the data and build much better models :)", "421276": "Learnmower I tried some different architectures too and they stopped improving after 3 or 4 epoches. \n\nNow it seems I've found a relatively \"good candidate\" (at least which works better than the others I tried)", "421284": "Serigne sounds about right on my end, too. 3-4 epochs depending on the architecture. I'm definitely using \"good\" loosely here. In Toxic, I think we traded off stability for score, but it was the blending later that helped to stabilize the variance.", "421309": "> I'm definitely using \"good\" loosely here\n\nSame for me....\"good\" is too relative in this phase of the competition :)", "421710": "@serigne, very good score by just one model.\nCan you please tell,\nhow many epochs are you training your single model **for each fold**?\n\nAnd 1200s in total or per fold?\n\nThanks.", "422099": "prashantkikani about 8 or 9 epoches for each fold \n\n1200s was remaining time after all training . \n\nActually I miscalculated. The entire model ( all folds) takes about 5800s. So it remains 1400s" }, "source": "meta" }