notebook.community

Edit and run



In [20]:

    
%reload_ext autoreload
%autoreload 2
%matplotlib inline

import matplotlib.pyplot as plt
import seaborn as sns
sns.set(context="notebook", style="white")

import numpy as np
import pandas as pd
import scipy.io as sio

import sys
sys.path.append('..')

from helper import general
from helper import pca



In [21]:

    
mat = sio.loadmat('./data/ex7faces.mat')
mat.keys()









    Out[21]:





dict_keys(['__header__', 'X', '__globals__', '__version__'])



In [22]:

    
X = np.array([x.reshape((32, 32)).T.reshape(1024) for x in mat.get('X')])
X.shape









    Out[22]:





(5000, 1024)

you do this to rotate the face clockwise 90 degree, I don't know why every pic I got from .mat file is in wrong orientation



In [23]:

    
pca.plot_n_image(X, n=64)

run PCA, find principle components



In [24]:

    
U, _, _ = pca.pca(X)



In [25]:

    
U.shape









    Out[25]:





(1024, 1024)



In [35]:

    
# didn't see face in principle components 
pca.plot_n_image(U, n=36)

I don't see face here...

reduce dimension to `k=100`



In [36]:

    
# no face in projected dimension too
Z = pca.project_data(X, U, k=100)
pca.plot_n_image(Z, n=64)

recover from `k=100`

you lost some detail, but.. they are strikingly similar



In [28]:

    
X_recover = pca.recover_data(Z, U)
pca.plot_n_image(X_recover, n=64)

sklearn PCA



In [29]:

    
from sklearn.decomposition import PCA



In [30]:

    
sk_pca = PCA(n_components=100)



In [31]:

    
Z = sk_pca.fit_transform(X)
Z.shape









    Out[31]:





(5000, 100)



In [32]:

    
pca.plot_n_image(Z, 64)



In [33]:

    
X_recover = sk_pca.inverse_transform(Z)
X_recover.shape









    Out[33]:





(5000, 1024)



In [34]:

    
pca.plot_n_image(X_recover, n=64)



In [ ]:

run PCA, find principle components

reduce dimension to k=100

recover from k=100

sklearn PCA

reduce dimension to `k=100`

recover from `k=100`