{ "id": 473178, "title": "Need help in identifying issue", "url": "/competitions/hms-harmful-brain-activity-classification/discussion/473178", "author_name": "Anmol Garg", "post_date": "2024-02-03T19:25:20.262000", "votes": 3, "comment_count": 2, "views": 0, "content": "

I was trying to train a wavenet based architecture on raw EEG data . After doing 5 Fold training I selected models having least respective fold validation loss(In range of 0.28 to 0.35) but after submitting with same model , LB score is coming out to be 0.67 I know that there is overfitting happening but isn't overfitting is when training loss decreases but not validation loss . What can I do to resolve the above . I am attaching an image from taining notebook and also the notebook link. Thanks. https://www.kaggle.com/code/anmolgarg1998/fork-of-hms-wavenet-inference-dfa251.\"img1\"\"img2\"

", "messages": [ { "id": 2634499, "postDate": "2024-02-03T19:25:20.263Z", "content": "

I was trying to train a wavenet based architecture on raw EEG data . After doing 5 Fold training I selected models having least respective fold validation loss(In range of 0.28 to 0.35) but after submitting with same model , LB score is coming out to be 0.67 I know that there is overfitting happening but isn't overfitting is when training loss decreases but not validation loss . What can I do to resolve the above . I am attaching an image from taining notebook and also the notebook link. Thanks. https://www.kaggle.com/code/anmolgarg1998/fork-of-hms-wavenet-inference-dfa251.\"img1\"\"img2\"

", "rawMarkdown": "I was trying to train a wavenet based architecture on raw EEG data . After doing 5 Fold training I selected models having least respective fold validation loss(In range of 0.28 to 0.35) but after submitting with same model , LB score is coming out to be 0.67 I know that there is overfitting happening but isn't overfitting is when training loss decreases but not validation loss . What can I do to resolve the above . I am attaching an image from taining notebook and also the notebook link. Thanks. [https://www.kaggle.com/code/anmolgarg1998/fork-of-hms-wavenet-inference-dfa251](url).![img1](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4463386%2F8cc362fa383bdf239c18290b74069da7%2Fval_loss.PNG?generation=1706988293309884&alt=media)![img2](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4463386%2F23edd77f83d0f824c71a9eab743ef74a%2Fval_loss2.PNG?generation=1706988309426195&alt=media)", "votes": 3 }, { "id": 2635919, "postDate": "2024-02-04T17:32:30.223Z", "content": "

I would check if there is any data leak that could be compromising your validation score…
\nAnother point that could be is to check the type of data separation you are doing, the leak could be there, the majority of the competition is grouping by patient_id

", "rawMarkdown": "I would check if there is any data leak that could be compromising your validation score...\nAnother point that could be is to check the type of data separation you are doing, the leak could be there, the majority of the competition is grouping by patient_id", "votes": 1 }, { "id": 2634947, "postDate": "2024-02-04T05:41:08.793Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2635919, "author_name": "Rafael Zimmermann", "author_url": "", "post_date": "2024-02-04T17:32:30.223000", "content": "

I would check if there is any data leak that could be compromising your validation score…
\nAnother point that could be is to check the type of data separation you are doing, the leak could be there, the majority of the competition is grouping by patient_id

", "votes": 1, "replies": [] }, { "id": 2634947, "author_name": "", "author_url": "", "post_date": "2024-02-04T05:41:08.793000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2634499": "I was trying to train a wavenet based architecture on raw EEG data . After doing 5 Fold training I selected models having least respective fold validation loss(In range of 0.28 to 0.35) but after submitting with same model , LB score is coming out to be 0.67 I know that there is overfitting happening but isn't overfitting is when training loss decreases but not validation loss . What can I do to resolve the above . I am attaching an image from taining notebook and also the notebook link. Thanks. [https://www.kaggle.com/code/anmolgarg1998/fork-of-hms-wavenet-inference-dfa251](url).![img1](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4463386%2F8cc362fa383bdf239c18290b74069da7%2Fval_loss.PNG?generation=1706988293309884&alt=media)![img2](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4463386%2F23edd77f83d0f824c71a9eab743ef74a%2Fval_loss2.PNG?generation=1706988309426195&alt=media)", "2635919": "I would check if there is any data leak that could be compromising your validation score...\nAnother point that could be is to check the type of data separation you are doing, the leak could be there, the majority of the competition is grouping by patient_id", "2634947": "" } }