My loss for my validation set while training is always around 0.04 or something similar and when I submit my results on gaggle the loss jumps to 1.9. I don't really understand why this happens, if someone could explain that would be good.
\n\nWhat exactly is wrong with the training? Is it overfitting?
", "messages": [ { "id": "166854", "postDate": "03/11/2017 14:06:24", "content": "My loss for my validation set while training is always around 0.04 or something similar and when I submit my results on gaggle the loss jumps to 1.9. I don't really understand why this happens, if someone could explain that would be good.
\n\nWhat exactly is wrong with the training? Is it overfitting?
", "rawMarkdown": "My loss for my validation set while training is always around 0.04 or something similar and when I submit my results on gaggle the loss jumps to 1.9. I don't really understand why this happens, if someone could explain that would be good. \n\nWhat exactly is wrong with the training? Is it overfitting?", "votes": null }, { "id": "166856", "postDate": "03/11/2017 14:10:24", "content": "Do you use validation by drivers? \nIf you haven't try using leave one group out validation based on drivers - this will likely improve your validation process and will prevent your model from overfitting the training data
", "rawMarkdown": "Do you use validation by drivers? \nIf you haven't try using leave one group out validation based on drivers - this will likely improve your validation process and will prevent your model from overfitting the training data", "votes": null }, { "id": "166857", "postDate": "03/11/2017 14:16:35", "content": "No I haven't tried that. I will though; you mean k-fold validation? or you mean using drivers to make sure the data is randomized properly? \nBut overfitting is the probable reason for this right ?
", "rawMarkdown": "No I haven't tried that. I will though; you mean k-fold validation? or you mean using drivers to make sure the data is randomized properly? \nBut overfitting is the probable reason for this right ?", "votes": null }, { "id": "166860", "postDate": "03/11/2017 14:32:35", "content": "I mean using drivers to make sure the data is randomized properly. \nAnd yes overfitting is the probable reason for this.\nCheck out scikit-learn leave one group out for an easy implementation
", "rawMarkdown": "I mean using drivers to make sure the data is randomized properly. \nAnd yes overfitting is the probable reason for this.\nCheck out scikit-learn leave one group out for an easy implementation", "votes": null }, { "id": "175974", "postDate": "04/18/2017 10:17:08", "content": "I'm facing the same problem with you, do you solve this problem? Does Nathan's advice work?
", "rawMarkdown": "I'm facing the same problem with you, do you solve this problem? Does Nathan's advice work?", "votes": null }, { "id": "333175", "postDate": "05/24/2018 14:54:50", "content": "It's depend what kind of loss have you choose in your CNN - kaggle use logloss..
", "rawMarkdown": "It's depend what kind of loss have you choose in your CNN - kaggle use logloss..", "votes": null } ], "comments": [ { "id": 166856, "author_name": "interesting", "author_url": "", "post_date": "03/11/2017 14:10:24", "content": "Do you use validation by drivers? \nIf you haven't try using leave one group out validation based on drivers - this will likely improve your validation process and will prevent your model from overfitting the training data
", "votes": null, "replies": [] }, { "id": 166857, "author_name": "amyhass", "author_url": "", "post_date": "03/11/2017 14:16:35", "content": "No I haven't tried that. I will though; you mean k-fold validation? or you mean using drivers to make sure the data is randomized properly? \nBut overfitting is the probable reason for this right ?
", "votes": null, "replies": [ { "id": 166860, "author_name": "interesting", "author_url": "", "post_date": "03/11/2017 14:32:35", "content": "I mean using drivers to make sure the data is randomized properly. \nAnd yes overfitting is the probable reason for this.\nCheck out scikit-learn leave one group out for an easy implementation
", "votes": null, "replies": [] } ] }, { "id": 175974, "author_name": "justinho", "author_url": "", "post_date": "04/18/2017 10:17:08", "content": "I'm facing the same problem with you, do you solve this problem? Does Nathan's advice work?
", "votes": null, "replies": [] }, { "id": 333175, "author_name": "", "author_url": "", "post_date": "05/24/2018 14:54:50", "content": "It's depend what kind of loss have you choose in your CNN - kaggle use logloss..
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "166854": "My loss for my validation set while training is always around 0.04 or something similar and when I submit my results on gaggle the loss jumps to 1.9. I don't really understand why this happens, if someone could explain that would be good. \n\nWhat exactly is wrong with the training? Is it overfitting?", "166856": "Do you use validation by drivers? \nIf you haven't try using leave one group out validation based on drivers - this will likely improve your validation process and will prevent your model from overfitting the training data", "166857": "No I haven't tried that. I will though; you mean k-fold validation? or you mean using drivers to make sure the data is randomized properly? \nBut overfitting is the probable reason for this right ?", "166860": "I mean using drivers to make sure the data is randomized properly. \nAnd yes overfitting is the probable reason for this.\nCheck out scikit-learn leave one group out for an easy implementation", "175974": "I'm facing the same problem with you, do you solve this problem? Does Nathan's advice work?", "333175": "It's depend what kind of loss have you choose in your CNN - kaggle use logloss.." }, "source": "meta" }