![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2416443%2Fca3f3e8c6fc19019f7b93be06e13906a%2Foutput.jpg?generation=1561933274065687&alt=media\")
Hi, I made my kernel, got good accuracy, saved submission.csv file that looks just like the file that others with published kernels have. But i am getting a public score 0.000 what can be the reason ? Pleaase help
", "messages": [ { "id": "565384", "postDate": "06/30/2019 21:35:16", "content": "Hi, I made my kernel, got good accuracy, saved submission.csv file that looks just like the file that others with published kernels have. But i am getting a public score 0.000 what can be the reason ? Pleaase help
", "rawMarkdown": "Hi, I made my kernel, got good accuracy, saved submission.csv file that looks just like the file that others with published kernels have. But i am getting a public score 0.000 what can be the reason ? Pleaase help", "votes": null }, { "id": "565392", "postDate": "06/30/2019 22:07:53", "content": "You need to share some more details, at this level of generality it's kinda hard to help...
", "rawMarkdown": "You need to share some more details, at this level of generality it's kinda hard to help...", "votes": null }, { "id": "565397", "postDate": "06/30/2019 22:22:29", "content": "Hi Konrad, thanks for responding. I train my kernel on kaggle and use these commands \n\nsample = pd.read_csv(\"../input/aptos2019-blindness-detection/sample_submission.csv\")\nsample.diagnosis = predicted_class_indices.astype(int)\nsample.to_csv('submission.csv',index = False)\n`\nThis is an image of my output. But still i get 0.00 public score. I read at some places that it can be because of different data types of columns. but its not the case here. I compared it with a guy who made his kernel public and got 70% accuracy. So what might be the problem ?
The kernel runs on a different test set (public + private) in the background when you submit, are you accounting for that?
", "rawMarkdown": "The kernel runs on a different test set (public + private) in the background when you submit, are you accounting for that?", "votes": null }, { "id": "565416", "postDate": "06/30/2019 23:18:28", "content": "I am getting a 90% train and test accuracy on dataset. Even if it runs on different dataset, it should atleast get something right randomly .. it cant be 0.000 still right ? I was searching this issue and most of the guys who had this issue said that their column data types were different but i verified column data types too. id_code is object and diagnosis is int. I even forced converted these 2 columns to str and int respectively. So what might be wrong now ? I dont see much difference in other guys's work who published kernels
", "rawMarkdown": "I am getting a 90% train and test accuracy on dataset. Even if it runs on different dataset, it should atleast get something right randomly .. it cant be 0.000 still right ? I was searching this issue and most of the guys who had this issue said that their column data types were different but i verified column data types too. id_code is object and diagnosis is int. I even forced converted these 2 columns to str and int respectively. So what might be wrong now ? I dont see much difference in other guys's work who published kernels", "votes": null }, { "id": "565421", "postDate": "06/30/2019 23:34:39", "content": "I had this issue before because I was using rescaled images uploaded to a dataset both for train and test. I needed to rescale the test images on the kernel so that it can work for a different test set.
", "rawMarkdown": "I had this issue before because I was using rescaled images uploaded to a dataset both for train and test. I needed to rescale the test images on the kernel so that it can work for a different test set.", "votes": null }, { "id": "565425", "postDate": "06/30/2019 23:41:20", "content": "Oh maybe that can be an issue but then i wonder why they have given option to upload our own dataset then ? There are so many features of keras like flow_from_directory etc that can be used if data is preprocessed and in folders making task easy but i guess i will just have to do it hard way. It already takes hours to compile and commit and grade.
\n\nThanks for the advice though. If you know any way to reduce this long process overhead let me know.
", "rawMarkdown": "Oh maybe that can be an issue but then i wonder why they have given option to upload our own dataset then ? There are so many features of keras like flow_from_directory etc that can be used if data is preprocessed and in folders making task easy but i guess i will just have to do it hard way. It already takes hours to compile and commit and grade. \n\nThanks for the advice though. If you know any way to reduce this long process overhead let me know.", "votes": null }, { "id": "565430", "postDate": "06/30/2019 23:58:29", "content": "I use preprocessed train data from a dataset (rescaled to 128x128 or 256x256). Only the test data needs to be fully processed on the kernel.
", "rawMarkdown": "I use preprocessed train data from a dataset (rescaled to 128x128 or 256x256). Only the test data needs to be fully processed on the kernel.", "votes": null }, { "id": "565431", "postDate": "07/01/2019 00:16:52", "content": "I get your point. I will try that. Thanks.
", "rawMarkdown": "I get your point. I will try that. Thanks.", "votes": null } ], "comments": [ { "id": 565392, "author_name": "konradb", "author_url": "", "post_date": "06/30/2019 22:07:53", "content": "You need to share some more details, at this level of generality it's kinda hard to help...
", "votes": null, "replies": [ { "id": 565397, "author_name": "shayanalibhatti", "author_url": "", "post_date": "06/30/2019 22:22:29", "content": "Hi Konrad, thanks for responding. I train my kernel on kaggle and use these commands \n\nsample = pd.read_csv(\"../input/aptos2019-blindness-detection/sample_submission.csv\")\nsample.diagnosis = predicted_class_indices.astype(int)\nsample.to_csv('submission.csv',index = False)\n`\nThis is an image of my output. But still i get 0.00 public score. I read at some places that it can be because of different data types of columns. but its not the case here. I compared it with a guy who made his kernel public and got 70% accuracy. So what might be the problem ?
The kernel runs on a different test set (public + private) in the background when you submit, are you accounting for that?
", "votes": null, "replies": [] }, { "id": 565416, "author_name": "shayanalibhatti", "author_url": "", "post_date": "06/30/2019 23:18:28", "content": "I am getting a 90% train and test accuracy on dataset. Even if it runs on different dataset, it should atleast get something right randomly .. it cant be 0.000 still right ? I was searching this issue and most of the guys who had this issue said that their column data types were different but i verified column data types too. id_code is object and diagnosis is int. I even forced converted these 2 columns to str and int respectively. So what might be wrong now ? I dont see much difference in other guys's work who published kernels
", "votes": null, "replies": [] }, { "id": 565421, "author_name": "mnpinto", "author_url": "", "post_date": "06/30/2019 23:34:39", "content": "I had this issue before because I was using rescaled images uploaded to a dataset both for train and test. I needed to rescale the test images on the kernel so that it can work for a different test set.
", "votes": null, "replies": [] }, { "id": 565425, "author_name": "shayanalibhatti", "author_url": "", "post_date": "06/30/2019 23:41:20", "content": "Oh maybe that can be an issue but then i wonder why they have given option to upload our own dataset then ? There are so many features of keras like flow_from_directory etc that can be used if data is preprocessed and in folders making task easy but i guess i will just have to do it hard way. It already takes hours to compile and commit and grade.
\n\nThanks for the advice though. If you know any way to reduce this long process overhead let me know.
", "votes": null, "replies": [] }, { "id": 565430, "author_name": "mnpinto", "author_url": "", "post_date": "06/30/2019 23:58:29", "content": "I use preprocessed train data from a dataset (rescaled to 128x128 or 256x256). Only the test data needs to be fully processed on the kernel.
", "votes": null, "replies": [] }, { "id": 565431, "author_name": "shayanalibhatti", "author_url": "", "post_date": "07/01/2019 00:16:52", "content": "I get your point. I will try that. Thanks.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "565384": "Hi, I made my kernel, got good accuracy, saved submission.csv file that looks just like the file that others with published kernels have. But i am getting a public score 0.000 what can be the reason ? Pleaase help", "565392": "You need to share some more details, at this level of generality it's kinda hard to help...", "565397": "Hi Konrad, thanks for responding. I train my kernel on kaggle and use these commands \n`\nsample` = pd.read_csv(\"../input/aptos2019-blindness-detection/sample_submission.csv\")\nsample.diagnosis = predicted_class_indices.astype(int)\nsample.to_csv('submission.csv',index = False)\n`\nThis is an image of my output. But still i get 0.00 public score. I read at some places that it can be because of different data types of columns. but its not the case here. I compared it with a guy who made his kernel public and got 70% accuracy. So what might be the problem ?\n\n", "565410": "The kernel runs on a different test set (public + private) in the background when you submit, are you accounting for that?", "565416": "I am getting a 90% train and test accuracy on dataset. Even if it runs on different dataset, it should atleast get something right randomly .. it cant be 0.000 still right ? I was searching this issue and most of the guys who had this issue said that their column data types were different but i verified column data types too. id_code is object and diagnosis is int. I even forced converted these 2 columns to str and int respectively. So what might be wrong now ? I dont see much difference in other guys's work who published kernels", "565421": "I had this issue before because I was using rescaled images uploaded to a dataset both for train and test. I needed to rescale the test images on the kernel so that it can work for a different test set.", "565425": "Oh maybe that can be an issue but then i wonder why they have given option to upload our own dataset then ? There are so many features of keras like flow_from_directory etc that can be used if data is preprocessed and in folders making task easy but i guess i will just have to do it hard way. It already takes hours to compile and commit and grade. \n\nThanks for the advice though. If you know any way to reduce this long process overhead let me know.", "565430": "I use preprocessed train data from a dataset (rescaled to 128x128 or 256x256). Only the test data needs to be fully processed on the kernel.", "565431": "I get your point. I will try that. Thanks." }, "source": "meta" }