Understanding through implementation

The goal of this repository is to understand Data Science algorithms through implementation. It also includes a generic docker image with tensorflow GPU to avoid package conflicts.

In order to use the GPU version, the nvidia runtime has to be installed and linked to docker. For that you need to install their drivers, nvidia-docker2, and link some configurations. You can find more detais in the links at the end.

To run:

    ./run
    # Other options
    # docker-compose up -d
    # docker compose up --build
    # Build the docker image
    # docker build -t data_science .
    # Run the container
    # docker run -it -p 8888:8888 -v data_science_volume:/workspace --name data_science_container data_science bash
    # Access jupyter lab at the url

Dependencies

• Nvidia drivers
• Docker
• Docker compose (plugin)
• Nvidia Container Toolkit

Useful links

https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html https://hub.docker.com/r/nvidia/cuda https://github.com/nvidia/nvidia-docker/wiki/Installation-(version-2.0) https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/user-guide.html https://github.com/compose-spec/compose-spec/blob/master/deploy.md#devices https://www.tensorflow.org/install/pip#hardware_requirements

Useful commands

nvidia-smi
nvidia-smi -lms 100
gpustat -cp --watch

ToDo list

• Multivariate linear regression, also using the normal equation, and polynomial and exponential regression
• Exponential (or a variable function: log, etc) regression with a time difference and scaling, to see if it's possible to fit any function
• Simple neural network
• Random forest
• Transformer
• GAN
• Dimension converter (like a semi-perfect dimensionality reduction tecnique) - train a neural network to represent some 3D data (cube) in 2D (maybe using encoders), without any missing information. Is that possible? How much more data would it take? If any n-dimension function could be reduced to a (giant) 2D function, maybe then it would be easier to visualize data and come up with better techniques.
• OLS vs MLE (maximum likelihood estimation)
• Neural network with fourier series? Other optimization/approximation methods?