{ "id": 459699, "title": "🥉 146th place – 9 models 1h inference - Segmentation through UNet", "url": "/competitions/child-mind-institute-detect-sleep-states/discussion/459699", "author_name": "Kirderf", "post_date": "2023-12-06T09:20:13.516000", "votes": 11, "comment_count": 0, "views": 0, "content": "

Short description of the solution below:

\n

As I only used 1-3 days of time for modifying, training and tuning the solution stayed at 9 models 1h inference but could easily been more modified and with also more models for better score.

\n

For the problem I used the base code from

\n\n

The original code had resnet34 as default and that was also the public pretrained model. Based on that I trained several of different models with Spec2DCNN: Segmentation through UNet - resnet18, efficientnet-b0 and resnext101_32x4d.

\n

Then I created a validation notebook below to get the validation score for some of the pretrained public and also the own trained models.

\n\n

The final inference solution was a torch stacked mean ensemble of 9 models using 1h test time inference with the same duration used, code below.

\n\n

With more time following had been tested: different architectures, more models, augmentation, using different durations…

\n

That’s it!
\nHappy Kaggling!

", "messages": [ { "id": 2550789, "postDate": "2023-12-06T09:20:13.517Z", "content": "

Short description of the solution below:

\n

As I only used 1-3 days of time for modifying, training and tuning the solution stayed at 9 models 1h inference but could easily been more modified and with also more models for better score.

\n

For the problem I used the base code from

\n\n

The original code had resnet34 as default and that was also the public pretrained model. Based on that I trained several of different models with Spec2DCNN: Segmentation through UNet - resnet18, efficientnet-b0 and resnext101_32x4d.

\n

Then I created a validation notebook below to get the validation score for some of the pretrained public and also the own trained models.

\n\n

The final inference solution was a torch stacked mean ensemble of 9 models using 1h test time inference with the same duration used, code below.

\n\n

With more time following had been tested: different architectures, more models, augmentation, using different durations…

\n

That’s it!
\nHappy Kaggling!

", "rawMarkdown": "Short description of the solution below:\n\nAs I only used 1-3 days of time for modifying, training and tuning the solution stayed at 9 models 1h inference but could easily been more modified and with also more models for better score.\n\nFor the problem I used the base code from\n\n- https://www.kaggle.com/code/tubotubo/cmi-submit\n- https://github.com/tubo213/kaggle-child-mind-institute-detect-sleep-states\n- https://www.kaggle.com/code/tubotubo/cmi-code\n\nThe original code had resnet34 as default and that was also the public pretrained model. Based on that I trained several of different models with Spec2DCNN: Segmentation through UNet - resnet18, efficientnet-b0 and resnext101_32x4d.\n\nThen I created a validation notebook below to get the validation score for some of the pretrained public and also the own trained models.\n\n- https://www.kaggle.com/kirderf/eval-child-mind-institute-detect-sleep-states\n\nThe final inference solution was a torch stacked mean ensemble of 9 models using 1h test time inference with the same duration used, code below.\n\n- https://www.kaggle.com/kirderf/cmi-submit-9-models-1h-inference.\n\nWith more time following had been tested: different architectures, more models, augmentation, using different durations…\n\nThat’s it!\nHappy Kaggling!\n", "votes": 11 } ], "comments": [], "raw_markdown_by_id": { "2550789": "Short description of the solution below:\n\nAs I only used 1-3 days of time for modifying, training and tuning the solution stayed at 9 models 1h inference but could easily been more modified and with also more models for better score.\n\nFor the problem I used the base code from\n\n- https://www.kaggle.com/code/tubotubo/cmi-submit\n- https://github.com/tubo213/kaggle-child-mind-institute-detect-sleep-states\n- https://www.kaggle.com/code/tubotubo/cmi-code\n\nThe original code had resnet34 as default and that was also the public pretrained model. Based on that I trained several of different models with Spec2DCNN: Segmentation through UNet - resnet18, efficientnet-b0 and resnext101_32x4d.\n\nThen I created a validation notebook below to get the validation score for some of the pretrained public and also the own trained models.\n\n- https://www.kaggle.com/kirderf/eval-child-mind-institute-detect-sleep-states\n\nThe final inference solution was a torch stacked mean ensemble of 9 models using 1h test time inference with the same duration used, code below.\n\n- https://www.kaggle.com/kirderf/cmi-submit-9-models-1h-inference.\n\nWith more time following had been tested: different architectures, more models, augmentation, using different durations…\n\nThat’s it!\nHappy Kaggling!\n" } }