this is the first time participating in deep learning competition :)\nis this competition hard to get to the top leaderboard without sufficient gpus on local?\nany advice would be helpful thanks
", "messages": [ { "id": "516009", "postDate": "04/13/2019 14:18:25", "content": "this is the first time participating in deep learning competition :)\nis this competition hard to get to the top leaderboard without sufficient gpus on local?\nany advice would be helpful thanks
", "rawMarkdown": "this is the first time participating in deep learning competition :)\nis this competition hard to get to the top leaderboard without sufficient gpus on local?\nany advice would be helpful thanks", "votes": null }, { "id": "516011", "postDate": "04/13/2019 14:18:54", "content": "FWIW, my current position is kernel only.
", "rawMarkdown": "FWIW, my current position is kernel only.", "votes": null }, { "id": "516051", "postDate": "04/13/2019 15:04:16", "content": "If you want to be the top, you need gpus, maybe you can borrow some from aws or gcp.
", "rawMarkdown": "If you want to be the top, you need gpus, maybe you can borrow some from aws or gcp.", "votes": null }, { "id": "516353", "postDate": "04/14/2019 03:44:22", "content": "thanks for the comment!
", "rawMarkdown": "thanks for the comment!", "votes": null }, { "id": "516354", "postDate": "04/14/2019 03:45:39", "content": "how many gpus are you using for this competition?? isn't it too expensive?
", "rawMarkdown": "how many gpus are you using for this competition?? isn't it too expensive?", "votes": null }, { "id": "517116", "postDate": "04/15/2019 15:06:49", "content": "I'm using 1080Ti + 1070 locally, but I'm also running multiple kernels on Kaggle.
", "rawMarkdown": "I'm using 1080Ti + 1070 locally, but I'm also running multiple kernels on Kaggle.", "votes": null }, { "id": "518674", "postDate": "04/17/2019 15:39:51", "content": "“You are permitted to train a model outside of Kernels and perform just the inference step from within Kernels.” In my opinion, you can use your local server or cloud server to train (powerful gpu 1080Ti or Titan XP), and get the model, then inference within Kernels to get your submission.csv. Relatively speaking, I don't know if it is a rigorous kernel game.
", "rawMarkdown": "“You are permitted to train a model outside of Kernels and perform just the inference step from within Kernels.” In my opinion, you can use your local server or cloud server to train (powerful gpu 1080Ti or Titan XP), and get the model, then inference within Kernels to get your submission.csv. Relatively speaking, I don't know if it is a rigorous kernel game.", "votes": null }, { "id": "523477", "postDate": "04/26/2019 11:35:36", "content": "I used only kernel and got LB 0.640. Now I start using local machine (1080ti x 3).
", "rawMarkdown": "I used only kernel and got LB 0.640. Now I start using local machine (1080ti x 3).", "votes": null }, { "id": "523489", "postDate": "04/26/2019 11:58:55", "content": "I'm also doing my experiment by executing multiple kernels. Though, the limitation is that we can't train deeper models with larger image size and batch size, but I think we can address those challenges using gradient accumulation and proper training strategies.
", "rawMarkdown": "I'm also doing my experiment by executing multiple kernels. Though, the limitation is that we can't train deeper models with larger image size and batch size, but I think we can address those challenges using gradient accumulation and proper training strategies.", "votes": null }, { "id": "523502", "postDate": "04/26/2019 12:23:39", "content": "You mean, got 0.640 by one kernel (in 9 hours) ? If so, that is amazing !
", "rawMarkdown": "You mean, got 0.640 by one kernel (in 9 hours) ? If so, that is amazing !", "votes": null }, { "id": "523512", "postDate": "04/26/2019 12:35:40", "content": "No. I used 7 kernels for train, then use one kernel to make prediction.
", "rawMarkdown": "No. I used 7 kernels for train, then use one kernel to make prediction.", "votes": null }, { "id": "523516", "postDate": "04/26/2019 12:42:14", "content": "Thanks! I'll try it.
", "rawMarkdown": "Thanks! I'll try it.", "votes": null }, { "id": "523518", "postDate": "04/26/2019 12:44:47", "content": "Hi，phalax，thank you for your sharing，I want to know if we can use multiple kernels to get results in the final.Is this allowed in kernel competition?
", "rawMarkdown": "Hi，phalax，thank you for your sharing，I want to know if we can use multiple kernels to get results in the final.Is this allowed in kernel competition?", "votes": null }, { "id": "523521", "postDate": "04/26/2019 12:50:25", "content": "You can see this in \"Kernels Requirements\".
\nSubmissions to this competition must be made through Kernels. You are permitted to train a model outside of Kernels and perform just the inference step from within Kernels.
I assume that the score is the result of ensembling 7 kernel outputs or single model, pls correct my assumption. Thanks!
", "rawMarkdown": "I assume that the score is the result of ensembling 7 kernel outputs or single model, pls correct my assumption. Thanks!", "votes": null }, { "id": "523527", "postDate": "04/26/2019 12:59:37", "content": "How to use multiple kernels ?
", "rawMarkdown": "How to use multiple kernels ?", "votes": null }, { "id": "523536", "postDate": "04/26/2019 13:13:38", "content": "Every models or experiment have each corresponding kernel and those kernels run simultaneously. You may refer on this query: https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52263
", "rawMarkdown": "Every models or experiment have each corresponding kernel and those kernels run simultaneously. You may refer on this query: https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52263", "votes": null } ], "comments": [ { "id": 516011, "author_name": "konradb", "author_url": "", "post_date": "04/13/2019 14:18:54", "content": "FWIW, my current position is kernel only.
", "votes": null, "replies": [ { "id": 516353, "author_name": "yangsaewon", "author_url": "", "post_date": "04/14/2019 03:44:22", "content": "thanks for the comment!
", "votes": null, "replies": [] } ] }, { "id": 516051, "author_name": "", "author_url": "", "post_date": "04/13/2019 15:04:16", "content": "If you want to be the top, you need gpus, maybe you can borrow some from aws or gcp.
", "votes": null, "replies": [ { "id": 516354, "author_name": "yangsaewon", "author_url": "", "post_date": "04/14/2019 03:45:39", "content": "how many gpus are you using for this competition?? isn't it too expensive?
", "votes": null, "replies": [] }, { "id": 517116, "author_name": "alexanderliao", "author_url": "", "post_date": "04/15/2019 15:06:49", "content": "I'm using 1080Ti + 1070 locally, but I'm also running multiple kernels on Kaggle.
", "votes": null, "replies": [] } ] }, { "id": 518674, "author_name": "suyukun", "author_url": "", "post_date": "04/17/2019 15:39:51", "content": "“You are permitted to train a model outside of Kernels and perform just the inference step from within Kernels.” In my opinion, you can use your local server or cloud server to train (powerful gpu 1080Ti or Titan XP), and get the model, then inference within Kernels to get your submission.csv. Relatively speaking, I don't know if it is a rigorous kernel game.
", "votes": null, "replies": [] }, { "id": 523477, "author_name": "phalanx", "author_url": "", "post_date": "04/26/2019 11:35:36", "content": "I used only kernel and got LB 0.640. Now I start using local machine (1080ti x 3).
", "votes": null, "replies": [ { "id": 523502, "author_name": "ttahara", "author_url": "", "post_date": "04/26/2019 12:23:39", "content": "You mean, got 0.640 by one kernel (in 9 hours) ? If so, that is amazing !
", "votes": null, "replies": [] }, { "id": 523512, "author_name": "phalanx", "author_url": "", "post_date": "04/26/2019 12:35:40", "content": "No. I used 7 kernels for train, then use one kernel to make prediction.
", "votes": null, "replies": [] }, { "id": 523516, "author_name": "ttahara", "author_url": "", "post_date": "04/26/2019 12:42:14", "content": "Thanks! I'll try it.
", "votes": null, "replies": [] }, { "id": 523518, "author_name": "hesene", "author_url": "", "post_date": "04/26/2019 12:44:47", "content": "Hi，phalax，thank you for your sharing，I want to know if we can use multiple kernels to get results in the final.Is this allowed in kernel competition?
", "votes": null, "replies": [] }, { "id": 523521, "author_name": "phalanx", "author_url": "", "post_date": "04/26/2019 12:50:25", "content": "You can see this in \"Kernels Requirements\".
\nSubmissions to this competition must be made through Kernels. You are permitted to train a model outside of Kernels and perform just the inference step from within Kernels.
I assume that the score is the result of ensembling 7 kernel outputs or single model, pls correct my assumption. Thanks!
", "votes": null, "replies": [] } ] }, { "id": 523489, "author_name": "projdev", "author_url": "", "post_date": "04/26/2019 11:58:55", "content": "I'm also doing my experiment by executing multiple kernels. Though, the limitation is that we can't train deeper models with larger image size and batch size, but I think we can address those challenges using gradient accumulation and proper training strategies.
", "votes": null, "replies": [ { "id": 523527, "author_name": "timmmmmms", "author_url": "", "post_date": "04/26/2019 12:59:37", "content": "How to use multiple kernels ?
", "votes": null, "replies": [] }, { "id": 523536, "author_name": "projdev", "author_url": "", "post_date": "04/26/2019 13:13:38", "content": "Every models or experiment have each corresponding kernel and those kernels run simultaneously. You may refer on this query: https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52263
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "516009": "this is the first time participating in deep learning competition :)\nis this competition hard to get to the top leaderboard without sufficient gpus on local?\nany advice would be helpful thanks", "516011": "FWIW, my current position is kernel only.", "516051": "If you want to be the top, you need gpus, maybe you can borrow some from aws or gcp.", "516353": "thanks for the comment!", "516354": "how many gpus are you using for this competition?? isn't it too expensive?", "517116": "I'm using 1080Ti + 1070 locally, but I'm also running multiple kernels on Kaggle.", "518674": "“You are permitted to train a model outside of Kernels and perform just the inference step from within Kernels.” In my opinion, you can use your local server or cloud server to train (powerful gpu 1080Ti or Titan XP), and get the model, then inference within Kernels to get your submission.csv. Relatively speaking, I don't know if it is a rigorous kernel game.", "523477": "I used only kernel and got LB 0.640. Now I start using local machine (1080ti x 3).", "523489": "I'm also doing my experiment by executing multiple kernels. Though, the limitation is that we can't train deeper models with larger image size and batch size, but I think we can address those challenges using gradient accumulation and proper training strategies.", "523502": "You mean, got 0.640 by one kernel (in 9 hours) ? If so, that is amazing !", "523512": "No. I used 7 kernels for train, then use one kernel to make prediction.", "523516": "Thanks! I'll try it.", "523518": "Hi，phalax，thank you for your sharing，I want to know if we can use multiple kernels to get results in the final.Is this allowed in kernel competition?", "523521": "You can see this in \"Kernels Requirements\".