Liverpool Ion Switching

In this repository you can find an outline of how to reproduce my 2nd place solution for Liverpool Ion Switching competition. It contains all the code and pipelines I used to create my winning submissions. If you run into any trouble with the setup/code or have any questions please contact me at st.dereka@gmail.com.

In ./summary.pdf you can find a detailed model description.

Post, which explains my approach to the challenge.

Kaggle kernel illustrating my preprocessing and data augmentation strategies.

Contents

• ./preprocessing - preprocessing scripts
• ./data - raw and preprocessed data
• ./config - configuration files (.json)
• ./models - serialized copies of models and their predictions
• ./model - train and inference pipelines of models
• ./postprocessing - code to write submissions and do some postprocessing
• ./submissions - final submissions
• ./evaluation - utilities to compute model's CV metrics

Software requirements

1. Python 3.6.9

2. CUDA 10.1

3. Nvidia Driver 418.67

4. Python packages are detailed in ./requirements.txt. In order to install them run:

    pip install -r requirements.txt
   

5. Ubuntu 18.04 LTS (it is not necessary to have exactly this OS installed, you can run the solution almost on any modern Linux distribution)

Hardware requirements (recommended)

These requirements should be fulfilled if you want to retrain all models from scratch. Running prediction with pretrained models consumes less resources - you don't even need a GPU.

1. 30 GB free disk space
2. 20 GB RAM
3. 1 x Tesla P100-PCIE-16GB
4. 1 x Intel Core i7-3720QM

Entry points

To make reproducing easier I created following scripts:

• ./prepare_data.py - reads parameters from ./config/PREPROCESSING.json and runs preprocessing pipeline
• ./train.py - reads parameters from ./config/RFC.json and ./config/WAVENET.json, runs training pipelines
• ./predict.py - reads parameters from ./config/RFC.json and ./config/WAVENET.json, runs inference pipelines and writes submissions.
• ./run_all_in_colab.ipynb - allows to reproduce all the results on Google Colab

How to reproduce the results?

The most simple way to reproduce the results is to run ./run_all_in_colab.ipynb on Google Colab. I prepared this entry point to make the process as simple as possible. If you want to reproduce the results on your local machine, follow these steps:

1. Clone the repo:

    git clone https://github.com/stdereka/liverpool-ion-switching.git
    cd liverpool-ion-switching
   

2. Download the data and pretrained models. Assumed, that you have Kaggle API installed and kaggle.json is generated and placed in appropriate directory. If you are the competition host, you can skip this step - all necessary data is in the package. Run:

    ./download_data.sh
   

3. Run preprocessing pipeline:

    python prepare_data.py
   

4. In order to reproduce two final submissions, run inference pipeline. Depending on your hardware, it will take about 10 minutes. Reproducing results is extremely simple: you don't even need a GPU. Two generated submissions in ./submissions directory reproduce my final LB score within a reasonable margin. Run following command:

    python predict.py --all
   

5. Retraining the models from scratch will take much more time and hardware resources. If you want to do it I suggest two options.

   1. Retrain only wavenet models on GPU (2nd layer of stacking). It takes 6-9 hours. Run:

        python train.py --wavenet
      

   2. Retrain all the models including RFCs and wavenets. It can take more than one day with hardware setup I described above. Run:

        python train.py --rfc --wavenet