{ "id": 452606, "title": "GRUNET and KLDivLoss", "url": "/competitions/child-mind-institute-detect-sleep-states/discussion/452606", "author_name": "", "post_date": "2023-11-02T18:54:00.766831300Z", "votes": 20, "comment_count": 3, "views": 0, "content": "

I've added a notebook that combines RNN and Convolution networks into a prediction for each timestep of the critical points.

\n

Train: https://www.kaggle.com/code/danielphalen/cmss-grunet-train
\nInfer: https://www.kaggle.com/code/danielphalen/cmss-grunet-infer

\n

I've also used KLDivLoss as the metric for how to train the model. Here is an example of how the training and predictions look:

\n

\"\"

\n

People should play with the architecture, there are improvements that can be made to smooth the outputs from transpose convolution.

", "messages": [ { "id": "2510163", "postDate": "11/02/2023 18:54:00", "content": "

I've added a notebook that combines RNN and Convolution networks into a prediction for each timestep of the critical points.

\n

Train: https://www.kaggle.com/code/danielphalen/cmss-grunet-train
\nInfer: https://www.kaggle.com/code/danielphalen/cmss-grunet-infer

\n

I've also used KLDivLoss as the metric for how to train the model. Here is an example of how the training and predictions look:

\n

\"\"

\n

People should play with the architecture, there are improvements that can be made to smooth the outputs from transpose convolution.

", "rawMarkdown": "I've added a notebook that combines RNN and Convolution networks into a prediction for each timestep of the critical points. \n\nTrain: [https://www.kaggle.com/code/danielphalen/cmss-grunet-train](https://www.kaggle.com/code/danielphalen/cmss-grunet-train)\nInfer: [https://www.kaggle.com/code/danielphalen/cmss-grunet-infer](https://www.kaggle.com/code/danielphalen/cmss-grunet-infer)\n\nI've also used KLDivLoss as the metric for how to train the model. Here is an example of how the training and predictions look:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5552115%2Faf9fc8953f9bdf32cf8c9bd722adb527%2Fnewplot.png?generation=1698951232264647&alt=media)\n\nPeople should play with the architecture, there are improvements that can be made to smooth the outputs from transpose convolution.", "votes": null }, { "id": "2510210", "postDate": "11/02/2023 19:36:25", "content": "

Great notebooks! Thanks for sharing!

", "rawMarkdown": "Great notebooks! Thanks for sharing!", "votes": null }, { "id": "2510561", "postDate": "11/03/2023 05:06:58", "content": "

Thanks for sharing.

", "rawMarkdown": "Thanks for sharing.", "votes": null }, { "id": "2510837", "postDate": "11/03/2023 09:45:45", "content": "

Nice page about KL Loss: https://github.com/christianversloot/machine-learning-articles/blob/main/how-to-use-kullback-leibler-divergence-kl-divergence-with-keras.md

", "rawMarkdown": "Nice page about KL Loss: https://github.com/christianversloot/machine-learning-articles/blob/main/how-to-use-kullback-leibler-divergence-kl-divergence-with-keras.md", "votes": null } ], "comments": [ { "id": 2510210, "author_name": "snnclsr", "author_url": "", "post_date": "11/02/2023 19:36:25", "content": "

Great notebooks! Thanks for sharing!

", "votes": null, "replies": [] }, { "id": 2510561, "author_name": "hookman", "author_url": "", "post_date": "11/03/2023 05:06:58", "content": "

Thanks for sharing.

", "votes": null, "replies": [] }, { "id": 2510837, "author_name": "atamazian", "author_url": "", "post_date": "11/03/2023 09:45:45", "content": "

Nice page about KL Loss: https://github.com/christianversloot/machine-learning-articles/blob/main/how-to-use-kullback-leibler-divergence-kl-divergence-with-keras.md

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2510163": "I've added a notebook that combines RNN and Convolution networks into a prediction for each timestep of the critical points. \n\nTrain: [https://www.kaggle.com/code/danielphalen/cmss-grunet-train](https://www.kaggle.com/code/danielphalen/cmss-grunet-train)\nInfer: [https://www.kaggle.com/code/danielphalen/cmss-grunet-infer](https://www.kaggle.com/code/danielphalen/cmss-grunet-infer)\n\nI've also used KLDivLoss as the metric for how to train the model. Here is an example of how the training and predictions look:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5552115%2Faf9fc8953f9bdf32cf8c9bd722adb527%2Fnewplot.png?generation=1698951232264647&alt=media)\n\nPeople should play with the architecture, there are improvements that can be made to smooth the outputs from transpose convolution.", "2510210": "Great notebooks! Thanks for sharing!", "2510561": "Thanks for sharing.", "2510837": "Nice page about KL Loss: https://github.com/christianversloot/machine-learning-articles/blob/main/how-to-use-kullback-leibler-divergence-kl-divergence-with-keras.md" }, "source": "meta" }