def reconstruction_error_classifier(training_image, train_label, face_image):
df = pd.DataFrame(training_image.T, index=train_label)
grouped = df.groupby(df.index)
min_e = float('inf')
min_label = df.index[0]
for key, tab in grouped:
# Compute Principle (eigen) subspace per class
train_image = tab.values.T
m, N = train_image.shape
eigvals, u_i = calc_eig_pca_small(train_image, m, N)
# Classification
face_mean = compute_avg_face(train_image)
phi = face_image - face_mean
w = np.dot(phi.T, u_i)
x_n = face_mean + np.dot(u_i, w)
e = np.linalg.norm(face_image - x_n)
if e < min_e:
min_label = key
min_e = e
return min_label
def commMachineRandDataXFeat(data, test_img, feature_list, N, m_lda):
pred = []
for k in data:
train_image = k[:, :-1].T
train_label = k[:, -1].T
m, N = train_image.shape
eigvals_pca, eigvecs_pca = calc_eig_pca_small(train_image, m, N)
face_avg = compute_avg_face(train_image)
phi_face = train_image - face_avg.reshape(face_avg.shape[0], 1)
# FLD
sw = compute_s_w(train_image, train_label, face_avg)
sb = compute_s_b(train_image, train_label, face_avg)
for l in feature_list:
m_eigvecs = eigvecs_pca[:, l]
eigvals_fld, eigvecs_fld, m_eigvecs = calc_eig_fld(
k_eigvecs=m_eigvecs, sw=sw, sb=sb)
## Variables
m_eigvecs_fld = eigvecs_fld[:, :m_lda]
w_opt = np.dot(m_eigvecs_fld.T, m_eigvecs.T).T.real
a_fld = np.dot(phi_face.T, w_opt)
out = list(
map(
lambda k: nn_classifier(a_fld, train_label, w_opt,
face_avg, k), test_img.T))
pred.append(out)
result_mat = np.array(pred)
result = stats.mode(result_mat, axis=0)
return result, result_mat
def calc_eig_fld(train_image=None,
train_label=None,
k=None,
sw=[],
sb=[],
k_eigvecs=[]):
if sw == [] or sb == [] or k_eigvecs == []:
m, N = train_image.shape
eigvals, eigvecs_pca = calc_eig_pca_small(train_image, m, N)
k_eigvecs = eigvecs_pca[:, :k]
face_avg = compute_avg_face(train_image)
sw = compute_s_w(train_image, train_label, face_avg)
sb = compute_s_b(train_image, train_label, face_avg)
rank_sw = np.linalg.matrix_rank(sw)
rank_sb = np.linalg.matrix_rank(sb)
# print("rank of sw: %a" %rank_sw)
# print("rank of sb: %a" %rank_sb)
wsww = np.dot(k_eigvecs.T, np.dot(sw, k_eigvecs))
wsbw = np.dot(k_eigvecs.T, np.dot(sb, k_eigvecs))
pca_mat = np.dot(np.linalg.pinv(wsww), wsbw)
eigvals_fld, eigvecs_fld = np.linalg.eig(pca_mat)
return eigvals_fld, eigvecs_fld, k_eigvecs
from data import train_image, train_label, test_image, test_label
from pca import calc_eig_pca_small
from fld import calc_eig_fld
from functions import compute_avg_face
from prediction import calc_accuracy, produce_pred_label, produce_reconstruction_pred_label
from ensemble import randFeature, bagging, commMachineRandDataAndFeat, commMachineRandDataXFeat
from sklearn.metrics import accuracy_score
import numpy as np
################################## Calculate accuracy #############################################
m, N = train_image.shape
k = 200
print(set(train_label))
# PCA
eigvals_pca, eigvecs_pca = calc_eig_pca_small(train_image, m, N)
## Variables
k_eigvecs = eigvecs_pca[:, :k]
face_avg = compute_avg_face(train_image)
phi_face = train_image - face_avg.reshape(m, 1)
a = np.dot(phi_face.T, k_eigvecs)
## Calcualte accuracy of pca
pred_label = produce_pred_label(a, train_label, k_eigvecs, face_avg,
test_image)
acc_pca = calc_accuracy(pred_label, test_label)
print(acc_pca)
## Calcualte accuracy of pca using reconstruction method
pred_label = produce_reconstruction_pred_label(train_image, train_label,