Because of the limit time of GPU, is there enough time to do ensemble?
", "messages": [ { "id": "504699", "postDate": "04/01/2019 03:28:22", "content": "Because of the limit time of GPU, is there enough time to do ensemble?
", "rawMarkdown": "Because of the limit time of GPU, is there enough time to do ensemble?", "votes": null }, { "id": "504862", "postDate": "04/01/2019 09:41:18", "content": "I believe for kaggler there are always enough time for ensembling but at this particular competition it might be better to spend time fintetuning your model.
", "rawMarkdown": "I believe for kaggler there are always enough time for ensembling but at this particular competition it might be better to spend time fintetuning your model.", "votes": null }, { "id": "504943", "postDate": "04/01/2019 10:57:13", "content": "Maybe snapshot ensembling can be usefull here (https://arxiv.org/pdf/1704.00109.pdf).
", "rawMarkdown": "Maybe snapshot ensembling can be usefull here (https://arxiv.org/pdf/1704.00109.pdf).", "votes": null }, { "id": "516004", "postDate": "04/13/2019 14:16:28", "content": "for this competition, do we have to ensemble models on local and inference on the kernel?? I'm not sure that I understood the competition right cause this is the first time for me to participate. thx :)
", "rawMarkdown": "for this competition, do we have to ensemble models on local and inference on the kernel?? I'm not sure that I understood the competition right cause this is the first time for me to participate. thx :)", "votes": null }, { "id": "516063", "postDate": "04/13/2019 15:13:52", "content": "First of all, it's not necessary to ensemble models - start with single one :) But yes, you train your models locally and then load your checkpoints as Datasets and do inference in kaggle kernel.
", "rawMarkdown": "First of all, it's not necessary to ensemble models - start with single one :) But yes, you train your models locally and then load your checkpoints as Datasets and do inference in kaggle kernel.", "votes": null }, { "id": "516358", "postDate": "04/14/2019 03:48:52", "content": "I thought ensemble was critical in deep learning competition. ensembling always(?) improves the model performance by little. isn't this right?? thanks I will start by one model first :)
", "rawMarkdown": "I thought ensemble was critical in deep learning competition. ensembling always(?) improves the model performance by little. isn't this right?? thanks I will start by one model first :)", "votes": null }, { "id": "516393", "postDate": "04/14/2019 04:58:40", "content": "What is special about this model, why is the snapshot integration method easy to use?
", "rawMarkdown": "What is special about this model, why is the snapshot integration method easy to use?", "votes": null }, { "id": "516395", "postDate": "04/14/2019 05:04:31", "content": "I used to did ensembling, but my score didn't boost, maybe my methods have problem, I will focus on it before the competition deadline
", "rawMarkdown": "I used to did ensembling, but my score didn't boost, maybe my methods have problem, I will focus on it before the competition deadline", "votes": null }, { "id": "516536", "postDate": "04/14/2019 11:29:12", "content": "will be helpful if you discuss about it later keep going!
", "rawMarkdown": "will be helpful if you discuss about it later keep going!", "votes": null }, { "id": "516623", "postDate": "04/14/2019 15:11:36", "content": "ok
", "rawMarkdown": "ok", "votes": null } ], "comments": [ { "id": 504862, "author_name": "yaroshevskiy", "author_url": "", "post_date": "04/01/2019 09:41:18", "content": "I believe for kaggler there are always enough time for ensembling but at this particular competition it might be better to spend time fintetuning your model.
", "votes": null, "replies": [ { "id": 516004, "author_name": "yangsaewon", "author_url": "", "post_date": "04/13/2019 14:16:28", "content": "for this competition, do we have to ensemble models on local and inference on the kernel?? I'm not sure that I understood the competition right cause this is the first time for me to participate. thx :)
", "votes": null, "replies": [] }, { "id": 516063, "author_name": "yaroshevskiy", "author_url": "", "post_date": "04/13/2019 15:13:52", "content": "First of all, it's not necessary to ensemble models - start with single one :) But yes, you train your models locally and then load your checkpoints as Datasets and do inference in kaggle kernel.
", "votes": null, "replies": [] }, { "id": 516358, "author_name": "yangsaewon", "author_url": "", "post_date": "04/14/2019 03:48:52", "content": "I thought ensemble was critical in deep learning competition. ensembling always(?) improves the model performance by little. isn't this right?? thanks I will start by one model first :)
", "votes": null, "replies": [] } ] }, { "id": 504943, "author_name": "mnpinto", "author_url": "", "post_date": "04/01/2019 10:57:13", "content": "Maybe snapshot ensembling can be usefull here (https://arxiv.org/pdf/1704.00109.pdf).
", "votes": null, "replies": [ { "id": 516393, "author_name": "", "author_url": "", "post_date": "04/14/2019 04:58:40", "content": "What is special about this model, why is the snapshot integration method easy to use?
", "votes": null, "replies": [] } ] }, { "id": 516395, "author_name": "", "author_url": "", "post_date": "04/14/2019 05:04:31", "content": "I used to did ensembling, but my score didn't boost, maybe my methods have problem, I will focus on it before the competition deadline
", "votes": null, "replies": [ { "id": 516536, "author_name": "yangsaewon", "author_url": "", "post_date": "04/14/2019 11:29:12", "content": "will be helpful if you discuss about it later keep going!
", "votes": null, "replies": [] }, { "id": 516623, "author_name": "", "author_url": "", "post_date": "04/14/2019 15:11:36", "content": "ok
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "504699": "Because of the limit time of GPU, is there enough time to do ensemble?", "504862": "I believe for kaggler there are always enough time for ensembling but at this particular competition it might be better to spend time fintetuning your model.", "504943": "Maybe snapshot ensembling can be usefull here (https://arxiv.org/pdf/1704.00109.pdf).", "516004": "for this competition, do we have to ensemble models on local and inference on the kernel?? I'm not sure that I understood the competition right cause this is the first time for me to participate. thx :)", "516063": "First of all, it's not necessary to ensemble models - start with single one :) But yes, you train your models locally and then load your checkpoints as Datasets and do inference in kaggle kernel.", "516358": "I thought ensemble was critical in deep learning competition. ensembling always(?) improves the model performance by little. isn't this right?? thanks I will start by one model first :)", "516393": "What is special about this model, why is the snapshot integration method easy to use?", "516395": "I used to did ensembling, but my score didn't boost, maybe my methods have problem, I will focus on it before the competition deadline", "516536": "will be helpful if you discuss about it later keep going!", "516623": "ok" }, "source": "meta" }