![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2341748%2F76691e040be9f3eb7770e5b426aa2733%2FScreenshot_2019-10-25%20Google%20Colaboratory.png?generation=1572005199216782&alt=media\")
I am getting low loss in both train and validation set(around 0.23 in train set and 0.34 in the validation set) but getting a score around 2.11, 8.33 when I submit my predictions. Can someone please tell me how to fix it?
", "messages": [ { "id": 657598, "postDate": "2019-10-25T09:32:54.250Z", "content": "Did you confirm the index of submission file?\nIn my case, the index order of sample submissions are different from the order of predicting test data.
", "rawMarkdown": "Did you confirm the index of submission file?\nIn my case, the index order of sample submissions are different from the order of predicting test data.", "votes": 1, "replies": [ { "id": 657749, "postDate": "2019-10-25T12:04:59.323Z", "content": "I checked the indexing of the submission is correct
", "rawMarkdown": "I checked the indexing of the submission is correct", "votes": 1 } ] }, { "id": 657679, "postDate": "2019-10-25T11:06:35.020Z", "content": "Wow, .23 and .34 are extremely large loss even for training. I think random guessing should result in ln(2)~=.693. On the bright side, simply flipping your prediction for your submission which scores 8.33 (turn probability prediction x into 1-x) should yield better results :)
", "rawMarkdown": "Wow, .23 and .34 are extremely large loss even for training. I think random guessing should result in ln(2)~=.693. On the bright side, simply flipping your prediction for your submission which scores 8.33 (turn probability prediction x into 1-x) should yield better results :)", "votes": 2 }, { "id": 656986, "postDate": "2019-10-24T20:22:26.873Z", "content": "I am getting low loss in both train and validation set(around 0.23 in train set and 0.34 in the validation set) but getting a score around 2.11, 8.33 when I submit my predictions. Can someone please tell me how to fix it?
", "rawMarkdown": "I am getting low loss in both train and validation set(around 0.23 in train set and 0.34 in the validation set) but getting a score around 2.11, 8.33 when I submit my predictions. Can someone please tell me how to fix it? ", "votes": 2 }, { "id": 657638, "postDate": "2019-10-25T10:17:37.337Z", "content": "Did you forget to sigmoid your prediction?
", "rawMarkdown": "Did you forget to sigmoid your prediction?", "replies": [ { "id": 657751, "postDate": "2019-10-25T12:06:43.387Z", "content": "No I used sigmoid \nhere is my model:-\n
Some time ago, I had a similar problem. Im my case it happened because I have set shuffle=True even for the test set data loader.
", "rawMarkdown": "Some time ago, I had a similar problem. Im my case it happened because I have set shuffle=True even for the test set data loader.", "replies": [ { "id": 657567, "postDate": "2019-10-25T08:45:08.297Z", "content": "I am not shuffling my test set so do you know what should be the fix of this problem?
", "rawMarkdown": "I am not shuffling my test set so do you know what should be the fix of this problem?" }, { "id": 657660, "postDate": "2019-10-25T10:40:11.197Z", "content": "I suggest you to compare again your exact pipeline for your test set with the one of the train/val set. Then, also compare it to the pipelines available in the public Kernels on Kaggle. Maybe you will find a inconsistency there.
", "rawMarkdown": "I suggest you to compare again your exact pipeline for your test set with the one of the train/val set. Then, also compare it to the pipelines available in the public Kernels on Kaggle. Maybe you will find a inconsistency there." } ] }, { "id": 656988, "postDate": "2019-10-24T20:25:58.253Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 657392, "postDate": "2019-10-25T05:24:05.750Z", "content": "I am using Binary Cross-Entropy
", "rawMarkdown": "I am using Binary Cross-Entropy" }, { "id": 657610, "postDate": "2019-10-25T09:42:38.573Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 657598, "author_name": "Inoichan", "author_url": "", "post_date": "2019-10-25T09:32:54.250000", "content": "Did you confirm the index of submission file?\nIn my case, the index order of sample submissions are different from the order of predicting test data.
", "votes": 1, "replies": [ { "id": 657749, "author_name": "Susnato Dhar", "author_url": "", "post_date": "2019-10-25T12:04:59.323000", "content": "I checked the indexing of the submission is correct
", "votes": 1, "replies": [] } ] }, { "id": 657679, "author_name": "Nicholas Lyu", "author_url": "", "post_date": "2019-10-25T11:06:35.020000", "content": "Wow, .23 and .34 are extremely large loss even for training. I think random guessing should result in ln(2)~=.693. On the bright side, simply flipping your prediction for your submission which scores 8.33 (turn probability prediction x into 1-x) should yield better results :)
", "votes": 2, "replies": [] }, { "id": 657638, "author_name": "Qishen Ha", "author_url": "", "post_date": "2019-10-25T10:17:37.337000", "content": "Did you forget to sigmoid your prediction?
", "votes": 0, "replies": [ { "id": 657751, "author_name": "Susnato Dhar", "author_url": "", "post_date": "2019-10-25T12:06:43.387000", "content": "No I used sigmoid \nhere is my model:-\n
Some time ago, I had a similar problem. Im my case it happened because I have set shuffle=True even for the test set data loader.
", "votes": 0, "replies": [ { "id": 657567, "author_name": "Susnato Dhar", "author_url": "", "post_date": "2019-10-25T08:45:08.297000", "content": "I am not shuffling my test set so do you know what should be the fix of this problem?
", "votes": 0, "replies": [] }, { "id": 657660, "author_name": "Alexander Abstreiter", "author_url": "", "post_date": "2019-10-25T10:40:11.197000", "content": "I suggest you to compare again your exact pipeline for your test set with the one of the train/val set. Then, also compare it to the pipelines available in the public Kernels on Kaggle. Maybe you will find a inconsistency there.
", "votes": 0, "replies": [] } ] }, { "id": 656988, "author_name": "", "author_url": "", "post_date": "2019-10-24T20:25:58.253000", "content": "", "votes": 0, "replies": [ { "id": 657392, "author_name": "Susnato Dhar", "author_url": "", "post_date": "2019-10-25T05:24:05.750000", "content": "I am using Binary Cross-Entropy
", "votes": 0, "replies": [] }, { "id": 657610, "author_name": "", "author_url": "", "post_date": "2019-10-25T09:42:38.573000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "657598": "Did you confirm the index of submission file?\nIn my case, the index order of sample submissions are different from the order of predicting test data.", "657679": "Wow, .23 and .34 are extremely large loss even for training. I think random guessing should result in ln(2)~=.693. On the bright side, simply flipping your prediction for your submission which scores 8.33 (turn probability prediction x into 1-x) should yield better results :)", "656986": "I am getting low loss in both train and validation set(around 0.23 in train set and 0.34 in the validation set) but getting a score around 2.11, 8.33 when I submit my predictions. Can someone please tell me how to fix it? ", "657638": "Did you forget to sigmoid your prediction?", "657478": "Some time ago, I had a similar problem. Im my case it happened because I have set shuffle=True even for the test set data loader.", "656988": "" } }