Should I apply augmentation on validation set? In the last competition, because test set was different from train set and had much lower quality images, I applied augmentation on val set to acquire more realistic results. What should I do in this competition? Thanks!
", "messages": [ { "id": "736296", "postDate": "02/04/2020 03:20:46", "content": "Should I apply augmentation on validation set? In the last competition, because test set was different from train set and had much lower quality images, I applied augmentation on val set to acquire more realistic results. What should I do in this competition? Thanks!
", "rawMarkdown": "Should I apply augmentation on validation set? In the last competition, because test set was different from train set and had much lower quality images, I applied augmentation on val set to acquire more realistic results. What should I do in this competition? Thanks!", "votes": null }, { "id": "736318", "postDate": "02/04/2020 04:08:58", "content": "In general, we do not apply any kind of augmentations to validation set except normailzation(and sometimes horizontal flip)
", "rawMarkdown": "In general, we do not apply any kind of augmentations to validation set except normailzation(and sometimes horizontal flip)", "votes": null }, { "id": "736333", "postDate": "02/04/2020 04:31:01", "content": "I think you are talking about Peking University / baidu :)\nNice to see you here again in this competition. I Don't think the test set will be different here much like we had in peking University/ baidu competition..good luck :)
", "rawMarkdown": "I think you are talking about Peking University / baidu :)\nNice to see you here again in this competition. I Don't think the test set will be different here much like we had in peking University/ baidu competition..good luck :)", "votes": null }, { "id": "736350", "postDate": "02/04/2020 04:54:28", "content": "Haha That's excatly what I was talking about. I think I'll just go with val set without augmentation then. Good luck to you too!
", "rawMarkdown": "Haha That's excatly what I was talking about. I think I'll just go with val set without augmentation then. Good luck to you too!", "votes": null }, { "id": "736702", "postDate": "02/04/2020 13:28:01", "content": "With regard to the concept of augmentation, i.e. making the data set larger, we would tend to only increase the training set. On a validation set, we would be evaluating the result of different augmentation approaches.
", "rawMarkdown": "With regard to the concept of **augmentation**, i.e. making the data set larger, we would tend to only increase the training set. On a validation set, we would be evaluating the result of different augmentation approaches.", "votes": null }, { "id": "736892", "postDate": "02/04/2020 17:13:41", "content": "Theoretically the validation set in order to remain a good feedback point should remain untouched (except normalization and resize). Practically if your test set is different from the validation one, you should bring the validation as close as possible to the real test set (to have a good feedback of the model). \nIn this competition, I do not think the public set and private one will be very different from the training data so I would leave the validation untouched
", "rawMarkdown": "Theoretically the validation set in order to remain a good feedback point should remain untouched (except normalization and resize). Practically if your test set is different from the validation one, you should bring the validation as close as possible to the real test set (to have a good feedback of the model). \nIn this competition, I do not think the public set and private one will be very different from the training data so I would leave the validation untouched", "votes": null } ], "comments": [ { "id": 736318, "author_name": "bibek777", "author_url": "", "post_date": "02/04/2020 04:08:58", "content": "In general, we do not apply any kind of augmentations to validation set except normailzation(and sometimes horizontal flip)
", "votes": null, "replies": [] }, { "id": 736333, "author_name": "mobassir", "author_url": "", "post_date": "02/04/2020 04:31:01", "content": "I think you are talking about Peking University / baidu :)\nNice to see you here again in this competition. I Don't think the test set will be different here much like we had in peking University/ baidu competition..good luck :)
", "votes": null, "replies": [ { "id": 736350, "author_name": "tonychenxyz", "author_url": "", "post_date": "02/04/2020 04:54:28", "content": "Haha That's excatly what I was talking about. I think I'll just go with val set without augmentation then. Good luck to you too!
", "votes": null, "replies": [] } ] }, { "id": 736702, "author_name": "ipythonx", "author_url": "", "post_date": "02/04/2020 13:28:01", "content": "With regard to the concept of augmentation, i.e. making the data set larger, we would tend to only increase the training set. On a validation set, we would be evaluating the result of different augmentation approaches.
", "votes": null, "replies": [] }, { "id": 736892, "author_name": "vladvdv", "author_url": "", "post_date": "02/04/2020 17:13:41", "content": "Theoretically the validation set in order to remain a good feedback point should remain untouched (except normalization and resize). Practically if your test set is different from the validation one, you should bring the validation as close as possible to the real test set (to have a good feedback of the model). \nIn this competition, I do not think the public set and private one will be very different from the training data so I would leave the validation untouched
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "736296": "Should I apply augmentation on validation set? In the last competition, because test set was different from train set and had much lower quality images, I applied augmentation on val set to acquire more realistic results. What should I do in this competition? Thanks!", "736318": "In general, we do not apply any kind of augmentations to validation set except normailzation(and sometimes horizontal flip)", "736333": "I think you are talking about Peking University / baidu :)\nNice to see you here again in this competition. I Don't think the test set will be different here much like we had in peking University/ baidu competition..good luck :)", "736350": "Haha That's excatly what I was talking about. I think I'll just go with val set without augmentation then. Good luck to you too!", "736702": "With regard to the concept of **augmentation**, i.e. making the data set larger, we would tend to only increase the training set. On a validation set, we would be evaluating the result of different augmentation approaches.", "736892": "Theoretically the validation set in order to remain a good feedback point should remain untouched (except normalization and resize). Practically if your test set is different from the validation one, you should bring the validation as close as possible to the real test set (to have a good feedback of the model). \nIn this competition, I do not think the public set and private one will be very different from the training data so I would leave the validation untouched" }, "source": "meta" }