Eigenface를 이용한 차원 축소와 SVM을 이용한 분류

 

 

 

 

Eigenface를 이용한 차원 축소와 SVM을 이용한 분류¶

 

In [1]:

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt


np.random.seed(2021)

 

 

1. Data¶

 

 

1.1 Data Load¶

데이터는 sklearn.datasets의 fetch_lfw_people로 받을 수 있습니다.

 

In [2]:

from sklearn.datasets import fetch_lfw_people

faces = fetch_lfw_people(min_faces_per_person=70, resize=0.4)

 

In [3]:

data, target = faces["data"], faces["target"]

 

 

1.2 Data EDA¶

 

 

이미지의 height와 width를 확인하면 다음과 같습니다.

 

In [4]:

n_samples, h, w = faces.images.shape

 

In [5]:

n_samples, h, w

 

Out[5]:

(1288, 50, 37)

 

 

얼굴의 주인들의 이름을 확인해 보겠습니다.

 

In [6]:

target_names = faces.target_names
n_classes = target_names.shape[0]

 

In [7]:

target_names

 

Out[7]:

array(['Ariel Sharon', 'Colin Powell', 'Donald Rumsfeld', 'George W Bush',
       'Gerhard Schroeder', 'Hugo Chavez', 'Tony Blair'], dtype='<U17')

 

 

이미지를 실제로 확인해 보겠습니다.

 

In [8]:

samples = data[:10].reshape(10, h, w)
fig, axes = plt.subplots(nrows=2, ncols=5, figsize=(20, 10))
for idx, sample in enumerate(samples):
    ax = axes[idx//5, idx%5]
    ax.imshow(sample, cmap="gray")
    ax.set_title(target_names[target[idx]])

 

 

 

 

1.3 Data Split¶

 

In [9]:

from sklearn.model_selection import train_test_split

train_data, test_data, train_target, test_target = train_test_split(
    data, target, train_size=0.7, random_state=2021
)

 

In [10]:

print(f"train_data size: {len(train_target)}, {len(train_target)/len(data):.2f}")
print(f"test_data size: {len(test_target)}, {len(test_target)/len(data):.2f}")

 

 

train_data size: 901, 0.70
test_data size: 387, 0.30

 

 

1.4 Data Scaling¶

 

In [11]:

from sklearn.preprocessing import StandardScaler

scaler = StandardScaler()

 

In [12]:

scaler.fit(train_data)

 

Out[12]:

StandardScaler(copy=True, with_mean=True, with_std=True)

 

In [13]:

scaled_train_data = scaler.transform(train_data)
scaled_test_data = scaler.transform(test_data)

 

 

2. Eigenface¶

 

 

Eigenface란 PCA를 이용해 얼굴 사진을 축소하면 생기는 eigenvector가 얼굴 모양과 같다고 하여서 생긴 용어입니다.
직접 실습을 통해 Eigenface를 생성해 보겠습니다.

 

 

2.1 학습¶

 

 

PCA를 이용해 데이터를 압축하겠습니다.

 

In [14]:

from sklearn.decomposition import PCA

pca = PCA()

 

In [15]:

pca.fit(scaled_train_data)

 

Out[15]:

PCA(copy=True, iterated_power='auto', n_components=None, random_state=None,
    svd_solver='auto', tol=0.0, whiten=False)

 

In [16]:

plt.plot(pca.explained_variance_ratio_.cumsum())
plt.axhline(0.9, color="red", linestyle="--")

 

Out[16]:

<matplotlib.lines.Line2D at 0x7f936116ced0>

 

 

 

explained variance ratio가 0.9가 되는 지점의 component를 사용하겠습니다.

 

In [17]:

pca = PCA(n_components=0.9)
pca.fit(scaled_train_data)

 

Out[17]:

PCA(copy=True, iterated_power='auto', n_components=0.9, random_state=None,
    svd_solver='auto', tol=0.0, whiten=False)

 

In [18]:

pca_train_data = pca.transform(scaled_train_data)
pca_test_data = pca.transform(scaled_test_data)

 

 

2.2 시각화¶

 

 

pca로 학습한 eigen vector를 시각화 해보겠습니다.
PCA를 통해 다음 eigen vector에 나오는 얼굴의 특징을 추출한다고 생각할 수 있습니다.

 

In [19]:

eigenfaces = pca.components_.reshape((pca.n_components_, h, w))
samples = eigenfaces[:10].reshape(10, h, w)
fig, axes = plt.subplots(nrows=2, ncols=5, figsize=(20, 10))
for idx, sample in enumerate(samples):
    ax = axes[idx//5, idx%5]
    ax.imshow(sample, cmap="gray")
    ax.set_title(f"eigenface {idx}")

 

 

 

 

3. SVM¶

 

 

3.1 Raw Data¶

우선 앞선 SVM 실습에서 진행했던 Baseline의 결과를 보겠습니다.

 

In [21]:

from sklearn.svm import SVC

svm = SVC()
svm.fit(scaled_train_data, train_target)

 

Out[21]:

SVC(C=1.0, break_ties=False, cache_size=200, class_weight=None, coef0=0.0,
    decision_function_shape='ovr', degree=3, gamma='scale', kernel='rbf',
    max_iter=-1, probability=False, random_state=None, shrinking=True,
    tol=0.001, verbose=False)

 

In [22]:

train_pred = svm.predict(scaled_train_data)
test_pred = svm.predict(scaled_test_data)

 

In [23]:

from sklearn.metrics import accuracy_score

train_acc = accuracy_score(train_target, train_pred)
test_acc = accuracy_score(test_target, test_pred)

 

In [24]:

print(f"train accuracy is {train_acc:.4f}")
print(f"test accuracy is {test_acc:.4f}")

 

 

train accuracy is 0.9567
test accuracy is 0.7339

 

 

3.2 Eigenface¶

이번에는 Eigenface로 추출된 특징만으로 SVM을 학습시킨 후 결과를 보겠습니다.

 

In [25]:

eigenface_svm = SVC()
eigenface_svm.fit(pca_train_data, train_target)

 

Out[25]:

SVC(C=1.0, break_ties=False, cache_size=200, class_weight=None, coef0=0.0,
    decision_function_shape='ovr', degree=3, gamma='scale', kernel='rbf',
    max_iter=-1, probability=False, random_state=None, shrinking=True,
    tol=0.001, verbose=False)

 

In [26]:

pca_train_pred = eigenface_svm.predict(pca_train_data)
pca_test_pred = eigenface_svm.predict(pca_test_data)

 

In [27]:

pca_train_acc = accuracy_score(train_target, pca_train_pred)
pca_test_acc = accuracy_score(test_target, pca_test_pred)

 

In [28]:

print(f"Eigenface train accuracy is {pca_train_acc:.4f}")
print(f"Eigenface test accuracy is {pca_test_acc:.4f}")

 

 

Eigenface train accuracy is 0.9390
Eigenface test accuracy is 0.7364

 

 

4. 마무리¶

 

In [29]:

train_data.shape

 

Out[29]:

(901, 1850)

 

In [30]:

pca_train_data.shape

 

Out[30]:

(901, 72)

 

In [31]:

print(f"Baseline test accuracy is {test_acc:.4f}")
print(f"Eigenface test accuracy is {pca_test_acc:.4f}")

 

 

Baseline test accuracy is 0.7339
Eigenface test accuracy is 0.7364

 

In [ ]: