PCA in pure Numpy (based on Sebastian Raschka tutorial)

pca_in_numpy.py

'''
Implementation of PCA with Numpy (using covariance), based on this tutorial by Sebastian Raschka: https://sebastianraschka.com/Articles/2014_pca_step_by_step.html
'''
import numpy as np 

def get_components(data: np.ndarray) -> np.ndarray:
    cov_mat = np.cov(data.T) # <-- get the covariance matrix

    ## calculate eigenvalues of the covariance matrix
    eig_val, eig_vec = np.linalg.eig(cov_mat)

    # sort components, largest to smallest
    idx_sort = np.flip(eig_val.argsort()) # <-- get ordering of eigenvectors: largest to smallest
    components = eig_vec[:,idx_sort]
    return components


## run example:
if __name__ == '__main__':

    ##__Generate Data
    data = np.random.normal(0,1,[100,3]) # <-- generates random data (assuming rows are instances, & columns are "features")

    ##__Get Components
    components = get_components(data)
    
    ##__Transform data using top 2 components (ie, matmul)
    num_components = 2
    transformed_data = data @ components[:,:num_components]