random-forests-cuda

This repo contains a random forest implementations from scratch using native python and a cuda GPU implemendation

Pleas see report.pdf for detailed description and timing analysis.

Running

• Clone repo
• Change directory to the root
• run python main.py

Overview of files in the repo

• main.py is an entry point that runs timing tests for fitting
• src/python/random_forest.py
  • class RandomForestFromScratch
    • Interface for running Random Forest models giving the same interface as sklearn Random Forests implementation
    • This class instantiates many different DecisionTreeNativePython or DecisionTreeCudaBase and trains them on random subsets of the training data.
    • When predicting the majority voting class of all the predictor trees are used
  • class DecisionTreeBase
    • Interface for the base Decision tree class giving same interface as sklearn Decision Tree Implemendation
    • Intended to subclass this and overwrite a few specific functions: calculate_split_scores, choose_best_score, split_data
    • This class includes the logic for iteratively training a decision tree by doing a breadth first training on the leaf nodes
    • Training is completed when the maximum depth is reached or leaf nodes only contain a single class
  • class DecisionTreeNativePython
    • calculate_split_scores naive serial implemendation
    • choose_best_score naive serial implemendation
    • split_data naive serial implemendation
  • class DecisionTreeCudaBase
    • calculate_split_scores points to src/utils/cuda_utils.DecisionTreeCudaUtils().split_scores
    • choose_best_score points to src/utils/cuda_utils.DecisionTreeCudaUtils().choose_best_score
    • split_data points to src/utils/cuda_utils.DecisionTreeCudaUtils().split_data
• src/python/utils/cuda_utils.py
  • class DecisionTreeCudaUtils
    • Intended to be the python interface into cuda functions
    • Written like the structure of our assignments
• src/cuda/kernels.cu
  • Cuda kernel functions written here
• src/python/utils.py
  • Includes generation of variable dimensional gaussian mixture model data for running different timing tests
  • Includes some timing objects to standardize timing across modules

Results

Results including figures and tabular data are shown in the results folder