I know this has been discussed in different threads, but I didn't find any given solution to the problem.
\nI am using this kernerl for inference. The only thing I changed is the model ( I am using EfficientNet ).
\nBut am I getting the \"Submission not found error\" when submitting.
Is there anyone who have encountered the same problem and fixed it ?
\nThanks in advance.
", "messages": [ { "id": 1044270, "postDate": "2020-10-09T16:33:24.027Z", "content": "I know this has been discussed in different threads, but I didn't find any given solution to the problem.
\nI am using this kernerl for inference. The only thing I changed is the model ( I am using EfficientNet ).
\nBut am I getting the \"Submission not found error\" when submitting.
Is there anyone who have encountered the same problem and fixed it ?
\nThanks in advance.
", "rawMarkdown": "I know this has been discussed in different threads, but I didn't find any given solution to the problem. \n\nI am using [this kernerl](https://www.kaggle.com/orkatz2/resnext-pulmonary-embolism-inference) for inference. The only thing I changed is the model ( I am using EfficientNet ). \nBut am I getting the \"Submission not found error\" when submitting. \n\nIs there anyone who have encountered the same problem and fixed it ?\n\nThanks in advance.", "votes": 4 }, { "id": 1044984, "postDate": "2020-10-10T09:08:03.580Z", "content": "Lung dataset returning None which in pytorch it is not supported to have None value in batches change it with empty list and check in prediction time for emptiness and skip that detection. For submission fill empty detections with zero and submit
", "rawMarkdown": "Lung dataset returning None which in pytorch it is not supported to have None value in batches change it with empty list and check in prediction time for emptiness and skip that detection. For submission fill empty detections with zero and submit", "votes": 1 }, { "id": 1044360, "postDate": "2020-10-09T17:34:25.257Z", "content": "Not sure with PyTorch, but with Keras you need to load EfficientNet and usually that assumes Internet Access. Try to run interactively without Internet Access and see what happens.
\nEfficientNet can be loaded from a Dataset, without Internet access.
", "rawMarkdown": "Not sure with PyTorch, but with Keras you need to load EfficientNet and usually that assumes Internet Access. Try to run interactively without Internet Access and see what happens. \n\nEfficientNet can be loaded from a Dataset, without Internet access.", "replies": [ { "id": 1044380, "postDate": "2020-10-09T17:52:54.290Z", "content": "Yeah that's what I did !
\nI troubleshooted the code and find out it's an OOM problem. I m trying to fix it.
Were you able to figure this out?
", "rawMarkdown": "Were you able to figure this out?" }, { "id": 1045197, "postDate": "2020-10-10T12:30:51.613Z", "content": "I discovered a huge reference count leak that I fixed by replacing: pred_df.append(mini_df) with pred_df.append(mini_df.values). However, I still see CPU RAM increasing too much.
Note: CPU RAM increases a lot on even on ResNeXt inference, which, even though it doesn't fail on a submission run, might fail on the final test set
", "rawMarkdown": "I discovered a huge reference count leak that I fixed by replacing: `pred_df.append(mini_df)` with `pred_df.append(mini_df.values)`. However, I still see CPU RAM increasing too much.\n\nNote: CPU RAM increases a lot on even on ResNeXt inference, which, even though it doesn't fail on a submission run, might fail on the final test set", "votes": 1 }, { "id": 1047608, "postDate": "2020-10-12T18:46:03.903Z", "content": "@philipkd The problem was with the index = 198 in the dataset, it contains more images than the rest batches, so when fitting it in memory we will have an OOM. I fixed this by creating a subdataloader.
", "rawMarkdown": "@philipkd The problem was with the index = 198 in the dataset, it contains more images than the rest batches, so when fitting it in memory we will have an OOM. I fixed this by creating a subdataloader.", "votes": 1 }, { "id": 1048900, "postDate": "2020-10-13T23:22:48.743Z", "content": "That's an incredible find, thanks!
", "rawMarkdown": "That's an incredible find, thanks!", "votes": 1 }, { "id": 1059782, "postDate": "2020-10-25T12:53:32.023Z", "content": "What is subdataloder. How did you fix it. I came across the same issue.. please help @nacirbouazizi
", "rawMarkdown": "What is subdataloder. How did you fix it. I came across the same issue.. please help @nacirbouazizi ", "votes": 1 }, { "id": 1059934, "postDate": "2020-10-25T15:50:41.210Z", "content": "@vrpirates You should create a small loader inside your big loader that can take half of the big batch size.
", "rawMarkdown": "@vrpirates You should create a small loader inside your big loader that can take half of the big batch size. \n" }, { "id": 1059954, "postDate": "2020-10-25T16:09:32.373Z", "content": "Can you share your code of sub dataloder…it will be helpful.
\nThank you
Lung dataset returning None which in pytorch it is not supported to have None value in batches change it with empty list and check in prediction time for emptiness and skip that detection. For submission fill empty detections with zero and submit
", "votes": 1, "replies": [] }, { "id": 1044360, "author_name": "quadcore/Richard Epstein", "author_url": "", "post_date": "2020-10-09T17:34:25.257000", "content": "Not sure with PyTorch, but with Keras you need to load EfficientNet and usually that assumes Internet Access. Try to run interactively without Internet Access and see what happens.
\nEfficientNet can be loaded from a Dataset, without Internet access.
", "votes": 0, "replies": [ { "id": 1044380, "author_name": "Nacir Bouazizi", "author_url": "", "post_date": "2020-10-09T17:52:54.290000", "content": "Yeah that's what I did !
\nI troubleshooted the code and find out it's an OOM problem. I m trying to fix it.
Were you able to figure this out?
", "votes": 0, "replies": [] }, { "id": 1045197, "author_name": "Philip Dhingra", "author_url": "", "post_date": "2020-10-10T12:30:51.613000", "content": "I discovered a huge reference count leak that I fixed by replacing: pred_df.append(mini_df) with pred_df.append(mini_df.values). However, I still see CPU RAM increasing too much.
Note: CPU RAM increases a lot on even on ResNeXt inference, which, even though it doesn't fail on a submission run, might fail on the final test set
", "votes": 1, "replies": [] }, { "id": 1047608, "author_name": "Nacir Bouazizi", "author_url": "", "post_date": "2020-10-12T18:46:03.903000", "content": "@philipkd The problem was with the index = 198 in the dataset, it contains more images than the rest batches, so when fitting it in memory we will have an OOM. I fixed this by creating a subdataloader.
", "votes": 1, "replies": [] }, { "id": 1048900, "author_name": "Philip Dhingra", "author_url": "", "post_date": "2020-10-13T23:22:48.743000", "content": "That's an incredible find, thanks!
", "votes": 1, "replies": [] }, { "id": 1059782, "author_name": "VR PIRATES", "author_url": "", "post_date": "2020-10-25T12:53:32.023000", "content": "What is subdataloder. How did you fix it. I came across the same issue.. please help @nacirbouazizi
", "votes": 1, "replies": [] }, { "id": 1059934, "author_name": "Nacir Bouazizi", "author_url": "", "post_date": "2020-10-25T15:50:41.210000", "content": "@vrpirates You should create a small loader inside your big loader that can take half of the big batch size.
", "votes": 0, "replies": [] }, { "id": 1059954, "author_name": "VR PIRATES", "author_url": "", "post_date": "2020-10-25T16:09:32.373000", "content": "Can you share your code of sub dataloder…it will be helpful.
\nThank you