In [1]:
%load_ext autoreload
%autoreload 2
%matplotlib inline
from numpy import *
from IPython.html.widgets import *
import matplotlib.pyplot as plt
from IPython.core.display import clear_output
PCA is a linear transformation. Suppose I have a $N \times P$ data matrix ${\bf X}$, where $N$ is the number of samples and $P$ is the dimension of each sample. Then PCA will find you a $K \times P$ matrix ${\bf V}$ such that
$$ \underbrace{{\bf X}}_{N \times P} = \underbrace{{\bf S}}_{P \times K} \underbrace{{\bf V}}_{K \times P}. $$Here, $K$ is the number of principal components with $K \le P$.
${\bf V}$ can be though of in many different ways.
The first way is to think of it as a de-correlating transformation: originally, each variable (or dimension) in ${\bf X}$ - there are $P$ of them - may be correlated. That is, if I take any two column vectors of ${\bf X}$, say ${\bf x}_0$ and ${\bf x}_1$, their covariance is not going to be zero.
Let's try this in a randomly generated data:
In [2]:
from numpy.random import standard_normal # Gaussian variables
N = 1000; P = 5
X = standard_normal((N, P))
W = X - X.mean(axis=0,keepdims=True)
print(dot(W[:,0], W[:,1]))
I'll skip ahead and use a pre-canned PCA routine from scikit-learn (but I'll dig into it a bit later!) Let's see what happens to the transformed variables, ${\bf S}$:
In [3]:
from sklearn.decomposition import PCA
S=PCA(whiten=True).fit_transform(X)
print(dot(S[:,0], S[:,1]))
Another way to look at ${\bf V}$ is to think of them as projections. Since the row vectors of ${\bf V}$ is orthogonal to each other, the projected data ${\bf S}$ lines in a new "coordinate system" specified by ${\bf V}$. Furthermore, the new coordinate system is sorted in the decreasing order of variance in the original data. So, PCA can be thought of as calculating a new coordinate system where the basis vectors point toward the direction of largest variances first.
Exercise 1. Let's get a feel for this in the following interactive example. Try moving the sliders around to generate the data, and see how the principal component vectors change.
In this demo, mu_x and mu_y specifies the center of the data, sigma_x and sigma_y the standard deviations, and everything is rotated by the angle theta. The two blue arrows are the rows of ${\bf V}$ that gets calculated.
When you click on center, the data is first centered (mean is subtracted from the data) first. (Question: why is it necessary to "center" data when mu_x and mu_y are not zero?)
In [4]:
from numpy.random import standard_normal
from matplotlib.patches import Ellipse
from numpy.linalg import svd
@interact
def plot_2d_pca(mu_x=FloatSlider(min=-3.0, max=3.0, value=0),
mu_y=FloatSlider(min=-3.0, max=3.0, value=0),
sigma_x=FloatSlider(min=0.2, max=1.8, value=1.8),
sigma_y=FloatSlider(min=0.2, max=1.8, value=0.3),
theta=FloatSlider(min=0.0, max=pi, value=pi/6), center=False):
mu=array([mu_x, mu_y])
sigma=array([sigma_x, sigma_y])
R=array([[cos(theta),-sin(theta)],[sin(theta),cos(theta)]])
X=dot(standard_normal((1000, 2)) * sigma[newaxis,:],R.T) + mu[newaxis,:]
# Plot the points and the ellipse
fig, ax = plt.subplots(figsize=(8,8))
ax.scatter(X[:200,0], X[:200,1], marker='.')
ax.grid()
M=8.0
ax.set_xlim([-M,M])
ax.set_ylim([-M,M])
e=Ellipse(xy=array([mu_x, mu_y]), width=sigma_x*3, height=sigma_y*3, angle=theta/pi*180,
facecolor=[1.0,0,0], alpha=0.3)
ax.add_artist(e)
# Perform PCA and plot the vectors
if center:
X_mean=X.mean(axis=0,keepdims=True)
else:
X_mean=zeros((1,2))
# Doing PCA here... I'm using svd instead of scikit-learn PCA, I'll come back to this.
U,s,V =svd(X-X_mean, full_matrices=False)
for v in dot(diag(s/sqrt(X.shape[0])),V): # Each eigenvector
ax.arrow(X_mean[0,0],X_mean[0,1],-v[0],-v[1],
head_width=0.5, head_length=0.5, fc='b', ec='b')
Ustd=U.std(axis=0)
ax.set_title('std(U*s) [%f,%f]' % (Ustd[0]*s[0],Ustd[1]*s[1]))
Yet another use for ${\bf V}$ is to perform a dimensionality reduction. In many scenarios you encounter in image manipulation (as I'll see soon), Imight want to have a more concise representation of the data ${\bf X}$. PCA with $K < P$ is one way to reduce the dimesionality: because PCA picks the directions with highest data variances, if a small number of top $K$ rows are sufficient to approximate (reconstruct) ${\bf X}$.
Well, we can use from sklearn.decomposition import PCA. But for learning, let's dig just one step into what it acutally does.
One of the easiest way to perform PCA is to use the singular value decomposition (SVD). SVD decomposes a matrix ${\bf X}$ into a unitary matrix ${\bf U}$, rectangular diagonal matrix ${\bf \Sigma}$ (called "singular values"), and another unitary matrix ${\bf W}$ such that
$$ {\bf X} = {\bf U} {\bf \Sigma} {\bf W}$$So how can Iuse that to do PCA? Well, it turns out ${\bf \Sigma} {\bf W}$ of SVD, are exactly what Ineed to calculate the ${\bf V}$ matrix for the PCA, so I just have to run SVD and set ${\bf V} = {\bf \Sigma} {\bf W}$.
(Note: svd of numpy returns only the diagonal elements of ${\bf \Sigma}$.)
Exercise 2. Generate 1000 10-dimensional data and perform PCA this way. Plot the squares of the singular values.
To reduce the the $P$-dimesional data ${\bf X}$ to a $K$-dimensional data, I just need to pick the top $K$ row vectors of ${\bf V}$ - let's call that ${\bf W}$ - then calcuate ${\bf T} = {\bf X} {\bf W}^\intercal$. ${\bf T}$ then has the dimension $N \times K$.
If I want to reconstruct the data ${\bf T}$, Isimply do ${\hat {\bf X}} = {\bf T} {\bf W}$ (and re-add the means for ${\bf X}$, if necessary).
Exercise 3. Reduce the same data to 5 dimensions, then based on the projected data ${\bf T}$, reconstruct ${\bf X}$. What's the mean squared error of the reconstruction?
Now that I have a handle on the PCA method, let's try applying it to a dataset consisting of face data. I will use the CAlifornia FAcial expressions dataset (CAFE) from http://cseweb.ucsd.edu/~gary/CAFE/ . The following code loads the dataset into the dataset variable:
In [5]:
import pickle
dataset=pickle.load(open('data/cafe.pkl','r'))
disp('dataset.images shape is %s' % str(dataset.images.shape))
disp('dataset.data shape is %s' % str(dataset.data.shape))
@interact
def plot_face(image_id=(0, dataset.images.shape[0]-1)):
plt.imshow(dataset.images[image_id],cmap='gray')
plt.title('Image Id = %d, Gender = %d' % (dataset.target[image_id], dataset.gender[image_id]))
plt.axis('off')
In [6]:
X=dataset.data.copy() # So that Iwon't mess up the data in the dataset\
X_mean=X.mean(axis=0,keepdims=True) # Mean for each dimension across sample (centering)
X_std=X.std(axis=0,keepdims=True)
X-=X_mean
disp(all(abs(X.mean(axis=0))<1e-12)) # Are means for all dimensions very close to zero?
Then I perform SVD to calculate the projection matrix $V$. By default, U,s,V=svd(...) returns full matrices, which will return $n \times n$ matrix U, $n$-dimensional vector of singular values s, and $d \times d$ matrix V. But here, I don't really need $d \times d$ matrix V; with full_matrices=False, svd only returns $n \times d$ matrix for V.
In [7]:
from numpy.linalg import svd
U,s,V=svd(X,compute_uv=True, full_matrices=False)
disp(str(U.shape))
disp(str(s.shape))
disp(str(V.shape))
I can also plot how much each eigenvector in V contributes to the overall variance by plotting variance_ratio = $\frac{s^2}{\sum s^2}$. (Notice that s is already in the decreasing order.) The cumsum (cumulative sum) of variance_ratio then shows how much of the variance is explained by components up to n_components.
In [8]:
variance_ratio=s**2/(s**2).sum() # Normalized so that they add to one.
@interact
def plot_variance_ratio(n_components=(1, len(variance_ratio))):
n=n_components-1
fig, axs = plt.subplots(1, 2, figsize=(12, 5))
axs[0].plot(variance_ratio)
axs[0].set_title('Explained Variance Ratio')
axs[0].set_xlabel('n_components')
axs[0].axvline(n, color='r', linestyle='--')
axs[0].axhline(variance_ratio[n], color='r', linestyle='--')
axs[1].plot(cumsum(variance_ratio))
axs[1].set_xlabel('n_components')
axs[1].set_title('Cumulative Sum')
captured=cumsum(variance_ratio)[n]
axs[1].axvline(n, color='r', linestyle='--')
axs[1].axhline(captured, color='r', linestyle='--')
axs[1].annotate(s='%f%% with %d components' % (captured * 100, n_components), xy=(n, captured),
xytext=(10, 0.5), arrowprops=dict(arrowstyle="->"))
Since I'm dealing with face data, each row vector of ${\bf V}$ is called an "eigenface". The first "eigenface" is the one that explains a lot of variances in the data, whereas the last one explains the least.
In [9]:
image_shape=dataset.images.shape[1:] # (H x W)
@interact
def plot_eigenface(eigenface=(0, V.shape[0]-1)):
v=V[eigenface]*X_std
plt.imshow(v.reshape(image_shape), cmap='gray')
plt.title('Eigenface %d (%f to %f)' % (eigenface, v.min(), v.max()))
plt.axis('off')
Now I'll try reconstructing faces with different number of principal components (PCs)! Now, the transformed X is reconstructed by multiplying by the sample standard deviations for each dimension and adding the sample mean. For this reason, even for zero components, you get a face-like image!
The rightmost plot is the "relative" reconstruction error (image minus the reconstruction squared, divided by the data standard deviations). White is where the error is close to zero, and black is where the relative error is large (1 or more). As you increase the number of PCs, you should see the error mostly going to zero (white).
In [10]:
@interact
def plot_reconstruction(image_id=(0,dataset.images.shape[0]-1), n_components=(0, V.shape[0]-1),
pc1_multiplier=FloatSlider(min=-2,max=2, value=1)):
# This is where Iperform the projection and un-projection
Vn=V[:n_components]
M=ones(n_components)
if n_components > 0:
M[0]=pc1_multiplier
X_hat=dot(multiply(dot(X[image_id], Vn.T), M), Vn)
# Un-center
I=X[image_id] + X_mean
I_hat = X_hat + X_mean
D=multiply(I-I_hat,I-I_hat) / multiply(X_std, X_std)
# And plot
fig, axs = plt.subplots(1, 3, figsize=(10, 10))
axs[0].imshow(I.reshape(image_shape), cmap='gray', vmin=0, vmax=1)
axs[0].axis('off')
axs[0].set_title('Original')
axs[1].imshow(I_hat.reshape(image_shape), cmap='gray', vmin=0, vmax=1)
axs[1].axis('off')
axs[1].set_title('Reconstruction')
axs[2].imshow(1-D.reshape(image_shape), cmap='gray', vmin=0, vmax=1)
axs[2].axis('off')
axs[2].set_title('Difference^2 (mean = %f)' % sqrt(D.mean()))
plt.tight_layout()
In [11]:
def plot_morph(left=0, right=1, mix=0.5):
# Projected images
x_lft=dot(X[left], V.T)
x_rgt=dot(X[right], V.T)
# Mix
x_avg = x_lft * (1.0-mix) + x_rgt * (mix)
# Un-project
X_hat = dot(x_avg[newaxis,:], V)
I_hat = X_hat + X_mean
# And plot
fig, axs = plt.subplots(1, 3, figsize=(10, 10))
axs[0].imshow(dataset.images[left], cmap='gray', vmin=0, vmax=1)
axs[0].axis('off')
axs[0].set_title('Left')
axs[1].imshow(I_hat.reshape(image_shape), cmap='gray', vmin=0, vmax=1)
axs[1].axis('off')
axs[1].set_title('Morphed (%.2f %% right)' % (mix * 100))
axs[2].imshow(dataset.images[right], cmap='gray', vmin=0, vmax=1)
axs[2].axis('off')
axs[2].set_title('Right')
plt.tight_layout()
interact(plot_morph,
left=IntSlider(max=dataset.images.shape[0]-1),
right=IntSlider(max=dataset.images.shape[0]-1,value=1),
mix=FloatSlider(value=0.5, min=0, max=1.0))
Out[11]:
(The answer: not very much...)
In [ ]: