nikhilmishradevelop/kaggle-child-mind-institute-detect-sleep-states

4th Place Solution(NikhilMishra's Part for Kaggle's competition Detection of Sleep States)

This is my part of 4th Place Solution for the Child Mind Institute - Detect Sleep States (Kaggle competition).

Detailed Document: https://www.kaggle.com/competitions/child-mind-institute-detect-sleep-states/discussion/459637

How to Reproduce (for Competition Organizers)

Hardware

• Cloud Service Used: https://cloud.jarvislabs.ai/ (RTX6000 Ada x 1)
• Cloud Service Framework: Tensorflow-2.12
• CPU: AMD EPYC 7713 64-Core Processor
• GPU: NVIDIA RTX 6000 Ada (48 GB)
• RAM: 128 GB

OS/platform

• Ubuntu 20.04.5 LTS

3rd-party software

• Python: 3.8.10
• Cuda: 12.2

Requirements

- Only the additional packages to be installed have been mentioned in requriements.txt (pip install -r requirements.txt)
- Rest of the packages are already present in the cloud service provider VM

Training

1. Upload competition dataset in /kaggle/input/child-mind-institute-detect-sleep-states

   • i.e. /kaggle/input/child-mind-institute-detect-sleep-states/train_series.parquet, etc... in the data folder as specified in settings.json
   • settings.json has the path specified relative to the notebooks

2. Run following notebook to prepare input dataset:

   • nbs/data_preprocessing.ipynb

3. Run follwing notebooks to train models:

   • nbs/fm_v13.ipynb
   • nbs/fm_v15.ipynb
   • nbs/fm_v20.ipynb
   • nbs/fm_v21.ipynb

4. ensembling_experiments.ipynb is notebook is not required to generate the final submission but you can use to see a basic ensembling example. A mix of manual and automatic hyperparmeter tuning was used to tune the weights for WBF.

NOTE:

• To avoid OOM, free memory when each notebook is finished executing by restarting the notebook or killing the kernel.

Inference(Predictions)

Please use this inference code for the final submitted solution https://www.kaggle.com/code/nikhilmishradev/fork-of-cmi-dss-final-sub-blend-new-postprocess

You can find the original competition inference code here https://www.kaggle.com/nikhilmishradev/cmi-dss-final-sub-blend-new-postprocess

• Both are the same (except the new one has retrained models to check reproducibility)
• Retrained models are marked as "-final" at the end

Supplemental Information for Competition Organizers

• Please use the cloud service provider VM environment for maximum reproducibility without any issues (refer #How to Reproduce)

• My models use tensorflow hence the results will differ slightly at each run, but the final scores on the test data will be almost the same.

• B7. Serialized copy of the trained models are in the 4 kaggle datasets (made public)

  Retrained Models

  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v13-final
  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v15-final
  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v20-final
  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v21-final

  Original Models

  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v13
  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v15
  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v20
  • https://www.kaggle.com/datasets/nikhilmishradev/sleep-model-fm-v21

• B8. entry_points.md is not included because my all codes are .ipynb format.