![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16438831%2F9e92c6880afd005e2634cea2ca6ad433%2Fpixel.png?generation=1710570938296012&alt=media\"){kind=tracking}
Hey Everyone,
\nHas anyone here tried training only on the 10 second snippets that were actually classified?
\nIf I understand correctly, the reviewers had the full 50 second EEGs and the full 10 minute spectrograms for context, but they were still tasked with classifying the event at the middle 10 seconds.
\nI've been wanting to experiment with this, but my time has been a little short and I haven't had a chance, so I thought I'd see if anyone's looked into this already.
\nI was thinking you could go about this in two ways.
\n1) Model only trained on these focused portion of the EEG &/or spectrogram.
\n2) Model trained on focused portion and whole EEG/spectrogram so that the whole portion gives context to the specific region of interest.
\nAs a related aside, has anyone here ever tried doing something like this: predict probabilities with model trained only on focused portion, then use the probabilities output by a model trained on the whole context to update the prediction similarly to how a Bayesian update would work?
", "messages": [ { "id": "2695859", "postDate": "03/13/2024 22:20:05", "content": "Hey Everyone,
\nHas anyone here tried training only on the 10 second snippets that were actually classified?
\nIf I understand correctly, the reviewers had the full 50 second EEGs and the full 10 minute spectrograms for context, but they were still tasked with classifying the event at the middle 10 seconds.
\nI've been wanting to experiment with this, but my time has been a little short and I haven't had a chance, so I thought I'd see if anyone's looked into this already.
\nI was thinking you could go about this in two ways.
\n1) Model only trained on these focused portion of the EEG &/or spectrogram.
\n2) Model trained on focused portion and whole EEG/spectrogram so that the whole portion gives context to the specific region of interest.
\nAs a related aside, has anyone here ever tried doing something like this: predict probabilities with model trained only on focused portion, then use the probabilities output by a model trained on the whole context to update the prediction similarly to how a Bayesian update would work?
", "rawMarkdown": "Hey Everyone,\n\nHas anyone here tried training only on the 10 second snippets that were actually classified? \n\nIf I understand correctly, the reviewers had the full 50 second EEGs and the full 10 minute spectrograms for context, but they were still tasked with classifying the event at the middle 10 seconds. \n\nI've been wanting to experiment with this, but my time has been a little short and I haven't had a chance, so I thought I'd see if anyone's looked into this already.\n\nI was thinking you could go about this in two ways.\n\n1) Model only trained on these focused portion of the EEG &/or spectrogram.\n\n2) Model trained on focused portion and whole EEG/spectrogram so that the whole portion gives context to the specific region of interest.\n\nAs a related aside, has anyone here ever tried doing something like this: predict probabilities with model trained only on focused portion, then use the probabilities output by a model trained on the whole context to update the prediction similarly to how a Bayesian update would work?", "votes": null }, { "id": "2695979", "postDate": "03/14/2024 01:26:26", "content": "I tried to use a hyperparameter to control the length of fragments and attempted to search for the optimal length by HPO. Based on @cdeotte's dataset, this method resulted in worse results in my experiments.
\nI think it may be due to the processing method of the dataset. In this dataset, each EEG data only generates one training sample. Therefore, cutting the middle 10 seconds may not be accurate.
\nSome public notebooks have used this method, you can refer to it:
\n", "rawMarkdown": "I tried to use a hyperparameter to control the length of fragments and attempted to search for the optimal length by HPO. Based on[ @cdeotte](https://www.kaggle.com/cdeotte)'s dataset, this method resulted in worse results in my experiments.\n\nI think it may be due to the processing method of the dataset. In this dataset, each EEG data only generates one training sample. Therefore, cutting the middle 10 seconds may not be accurate.\n\nSome public notebooks have used this method, you can refer to it: \n1. [HMS Resnet1d GRU Train - 1 / 5 Dataset](https://www.kaggle.com/code/konstantinboyko/hms-resnet1d-gru-train-1-5-dataset#Dataset)", "votes": null }, { "id": "2699804", "postDate": "03/16/2024 06:35:58", "content": "How about concatenating 10m & 50s & 20s?
\nExample below
This looks much more effective than only using center-cropping data. I will try it later.
", "rawMarkdown": "This looks much more effective than only using center-cropping data. I will try it later.", "votes": null } ], "comments": [ { "id": 2695979, "author_name": "zijiangyang1116", "author_url": "", "post_date": "03/14/2024 01:26:26", "content": "I tried to use a hyperparameter to control the length of fragments and attempted to search for the optimal length by HPO. Based on @cdeotte's dataset, this method resulted in worse results in my experiments.
\nI think it may be due to the processing method of the dataset. In this dataset, each EEG data only generates one training sample. Therefore, cutting the middle 10 seconds may not be accurate.
\nSome public notebooks have used this method, you can refer to it:
\n", "votes": null, "replies": [] }, { "id": 2699804, "author_name": "seoyunje", "author_url": "", "post_date": "03/16/2024 06:35:58", "content": "How about concatenating 10m & 50s & 20s?
\nExample below
This looks much more effective than only using center-cropping data. I will try it later.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2695859": "Hey Everyone,\n\nHas anyone here tried training only on the 10 second snippets that were actually classified? \n\nIf I understand correctly, the reviewers had the full 50 second EEGs and the full 10 minute spectrograms for context, but they were still tasked with classifying the event at the middle 10 seconds. \n\nI've been wanting to experiment with this, but my time has been a little short and I haven't had a chance, so I thought I'd see if anyone's looked into this already.\n\nI was thinking you could go about this in two ways.\n\n1) Model only trained on these focused portion of the EEG &/or spectrogram.\n\n2) Model trained on focused portion and whole EEG/spectrogram so that the whole portion gives context to the specific region of interest.\n\nAs a related aside, has anyone here ever tried doing something like this: predict probabilities with model trained only on focused portion, then use the probabilities output by a model trained on the whole context to update the prediction similarly to how a Bayesian update would work?", "2695979": "I tried to use a hyperparameter to control the length of fragments and attempted to search for the optimal length by HPO. Based on[ @cdeotte](https://www.kaggle.com/cdeotte)'s dataset, this method resulted in worse results in my experiments.\n\nI think it may be due to the processing method of the dataset. In this dataset, each EEG data only generates one training sample. Therefore, cutting the middle 10 seconds may not be accurate.\n\nSome public notebooks have used this method, you can refer to it: \n1. [HMS Resnet1d GRU Train - 1 / 5 Dataset](https://www.kaggle.com/code/konstantinboyko/hms-resnet1d-gru-train-1-5-dataset#Dataset)", "2699804": "How about concatenating 10m & 50s & 20s? \n\nExample below", "2700441": "This looks much more effective than only using center-cropping data. I will try it later." }, "source": "meta" }