problem 1: PCA

hw4/pca.py

@@ -0,0 +1,107 @@
+#!/usr/env/bin python
+import sys, os
+import argparse
+import numpy as np
+from scipy import misc
+import matplotlib.pyplot as plt
+
+def read_pic(file, obj_size=10, pic_size=10):
+ X = []
+ base = 65
+ global width, height
+ pic = None
+ for o in range(obj_size):
+ for p in range(pic_size):
+ pic_file = '{}{:02}.bmp'.format(chr(base+o), p)
+ # print(pic_file)
+ pic = misc.imread(os.path.join(pic_dir, pic_file))
+ X.append(pic.flatten())
+ width = pic.shape[0]
+ height = pic.shape[1]
+ return np.array(X)
+
+
+def save_img(data, filename='default', subplot=False, size=0):
+ global width, height
+
+ if subplot:
+ fig = plt.figure(figsize=(16, 16))
+ for i in range(size):
+ ax = fig.add_subplot(np.sqrt(size), np.sqrt(size), i+1)
+ ax.imshow(data[i].reshape(width, height), cmap='gray')
+ plt.xticks(np.array([]))
+ plt.yticks(np.array([]))
+ plt.tight_layout()
+ fig.savefig(filename)
+ else:
+ fig = plt.figure(figsize=(8, 8))
+ plt.imshow(data, cmap='gray')
+ plt.xticks(np.array([]))
+ plt.yticks(np.array([]))
+ plt.tight_layout()
+ fig.savefig(filename)
+
+
+def pca(x, face_size):
+ global width, height
+ mu = np.mean(x, axis=0)
+ X = x - mu
+
+ eigen_faces, sigma, v = np.linalg.svd(X.T, full_matrices=False)
+ picked_faces = eigen_faces.T[:face_size]
+ weights = np.dot(X, picked_faces.T)
+
+ return mu.reshape(width, height), weights, picked_faces
+
+
+def reconstruct(weights, eigen_faces, eigen_size=0):
+ pics = np.dot(weights, eigen_faces)
+ save_img(pics, filename='original_eigen_{}.png'.format(eigen_size),
+ subplot=True, size=pics.shape[0])
+
+def cal_error(X, error):
+ print('Start calculating...')
+ for i in range(1, 101):
+ mu, weights, eigen_faces = pca(X, i)
+ pics = np.dot(weights, eigen_faces)
+ rsme = np.sqrt(np.mean(np.square(X - mu.flatten() - pics))) / 256
+ # print('Now {}, RSME: {}'.format(i, rsme))
+ if rsme < error :
+ return i
+
+
+def main():
+ pics_matrix = read_pic(pic_dir)
+
+ mu, weights, eigen_faces = pca(pics_matrix, 9)
+
+ # Save mean face
+ #misc.imsave('average_face.png', mu)
+ save_img(mu, filename='average_face.png')
+
+ # Save eigen faces
+ save_img(eigen_faces, filename='eigen_faces.png',
+ subplot=True, size=eigen_faces.shape[0])
+
+ # Save original image
+ save_img(pics_matrix, filename='original.png',
+ subplot=True, size=pics_matrix.shape[0])
+
+ # Reconstrcut by eigen faces
+ mu, weights, eigen_faces = pca(pics_matrix, 5)
+ reconstruct(weights, eigen_faces, eigen_size=eigen_faces.shape[0])
+
+ # Calculate the RSME
+ min_size_of_eigen_faces = cal_error(pics_matrix, 0.01)
+ print('>>>{}<<<'.format(min_size_of_eigen_faces))
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(description='problem 1: PCA')
+ parser.add_argument('--data', metavar='<#data>', type=str, required=True)
+ args = parser.parse_args()
+
+ width = height = -1
+ pic_dir = './' + args.data
+
+ main()