Notes about the book The Kaggle book

• Chapter 1 - Introducing Kaggle and other data science competition
• Chapter 2 - Organizing Data with Datasets
  • Setting up a dataset
  • Gathering data
  • Working with datasets
  • Using Kaggle datasets in Google Colab
• Chapter 3 - Working and Learning with Kaggle Notebooks
  • Setting up a notebook
  • Running your notebook
  • Saving notebooks to GitHub
  • Getting the most out of notebooks
  • Kaggle Learn courses
• Chapter 4 - Leveraging Discussion Foruns

Chapter 1 - Introducing Kaggle and other data science competition

• Kaggle public API docs.
• Kaggle API Github repo.
• Tip: Interact with others on the discussion forum when enrolled on a competition to share and learn.
• Common Task Framework (CTF): Great for advancing state of the art solutions.
  • Well defined metrics and quality data
  • Competition
  • Sharing between competitors
  • Compute-resource availability
• What can go wrong in a competition:
  • Leakeage from the data: data contain informatio of the target not available in real-time.
  • Probing from the leaderboard: Use the leaderboard to metric to tune your solution.
    • Example: https://www.kaggle.com/c/dont-overfit-ii/discussion/91766
  • Overfitting and consequent leaderboard shake-up: cases with huge gap between the training set and the public test set
    • Technique to measure discrepancies between training set and test set: https://www.kaggle.com/code/tunguz/adversarial-ieee/notebook
  • Private sharing
• Jeremy Howard on how to set you up for success on Kaggle: https://www.kaggle.com/code/jhoward/first-steps-road-to-the-top-part-1

Chapter 2 - Organizing Data with Datasets

Setting up a dataset

• It is possible to upload a dataset either privately or publicly.
• It is possible to use "Import a GitHub repository" option to import a experimental library not yet available on Kaggle Notebooks.

Gathering data

• Interesting interview from Larxel:
  • On creating datasets:
    • All in all, the process that I recommend starts with setting your purpose, breaking it down into objectives and topics, formulating questions to fulfil these topics, surveying possible sources of data, selecting and gathering, pre-processing, documenting, publishing, maintaining and supporting, and finally, improvement actions.
  • On learning on Kaggle:
    • Absorbing all the knowledge at the end of a competition
    • Replication of winning solutions in finished competitions

Working with datasets

• The easiest way to work with Kaggle datasets is by creating a notebook from the dataset webpage.

Using Kaggle datasets in Google Colab

• This section contains a step-by-step to download Kaggle Datasets into Colab.

  1. Download Kaggle API from your Kaggle account. Place it ~/.kaggle/kaggle.json
  2. Create folder Kaggle on your GDrive and upload .json there.
  3. Mount GDrive to your colab

     from google.colab import drive
     drive.mount('/content/gdrive')

  4. Provide path to .json config

     import os
     # content/gdrive/My Drive/Kaggle is the path where kaggle.json is 
     # present in the Google Drive
     os.environ['KAGGLE_CONFIG_DIR'] = "/content/gdrive/My Drive/Kaggle"
     # change the working directory
     %cd /content/gdrive/My Drive/Kaggle
     

  5. Go to the dataset page and use the copy API command.
  6. Run the command on the colab. Ex.:

     !kaggle datasets download -d bricevergnou/spotify-recommendation

  7. Data is downloaded to os.environ['KAGGLE_CONFIG_DIR']. Unzip it and you are ready to go.

Chapter 3 - Working and Learning with Kaggle Notebooks

Setting up a notebook

• Create from scratch
• Create from a dataset
• Fork from a notebook.
  • Good manner to upvote the original notebook and give credit if building solution on top.

Running your notebook

• You can either run the notebook interactvely or save and run specific versions.

Saving notebooks to GitHub

• It is possible to enable automatic sync between Kaggle Notebooks and Github.
  • File -> Link to Github

Getting the most out of notebooks

• Use your free GPU/TPU wisely
  • The hours start counting as soon as you start your notebook.
  • Disable GPU until you write your code, check for syntax and run on small subset of data to check for errors.
  • Once it is ready to run change the run-time accelerator.
• Sometimes Kaggle notebooks are not enough. Example data > 100GB.
  • Downsample data -> affect performance.
  • Use Kaggle datasets on Google Colab as shown in chapter 2.
  • Upgrade your Kaggle notebook to use the Google Cloud AI Notebooks. It is not free.
• Nice interview from heads or tails, who specializes onn EDA.

Kaggle Learn courses

• Kaggle has a list of courses to learn about specific topics.
• Nice interview Andrada Olteanu about how to use notebooks to learn.

Chapter 4 - Leveraging Discussion Foruns

• Not really noteworthy in the chapter. Just use the discussion forum to share and learn from other Kagglers.