用keras的cnn做人臉分類

keras介紹

Keras是一個簡約，高度模塊化的神經網絡庫。采用Python / Theano開發。
使用Keras如果你需要一個深度學習庫：

可以很容易和快速實現原型（通過總模塊化，極簡主義，和可擴展性）
同時支持卷積網絡（vision）和復發性的網絡（序列數據）。以及兩者的組合。
無縫地運行在CPU和GPU上。
keras的資源庫網址為https://github.com/fchollet/keras

olivettifaces人臉數據庫介紹

Olivetti Faces是紐約大學的一個比較小的人臉庫，由 40個人的400張圖片構成，即每個人的人臉圖片為10張。每張圖片的灰度級為8位，每個像素的灰度大小位於0-255之間，每張圖片大小為64×64。 如下圖，這個圖片大小是1140942，一共有2020張人臉，故每張人臉大小是（1140/20）（942/20）即5747=2679：

預處理模塊

使用了PIL（Python Imaging Library）模塊，是Python平台事實上的圖像處理標准庫。
預處理流程是：打開文件-》歸一化-》將圖片轉為數據集-》生成label-》使用pickle序列化數據集

numpy.ndarray.flatten函數的功能是將一個矩陣平鋪為向量

from PIL import Image
import numpy
import cPickle

img = Image.open('G:\data\olivettifaces.gif')
# numpy supports conversion from image to ndarray and normalization by dividing 255
# 1140 * 942 ndarray
img_ndarray = numpy.asarray(img, dtype='float64') / 255
# create numpy array of 400*2679
img_rows, img_cols = 57, 47
face_data = numpy.empty((400, img_rows*img_cols))
# convert 1140*942 ndarray to 400*2679 matrix

for row in range(20):
    for col in range(20):
        face_data[row*20+col] = numpy.ndarray.flatten(img_ndarray[row*img_rows:(row+1)*img_rows, col*img_cols:(col+1)*img_cols])

# create label
face_label = numpy.empty(400, dtype=int)
for i in range(400):
    face_label[i] = i / 10

# pickling file
f = open('G:\data\olivettifaces.pkl','wb')
# store data and label as a tuple
cPickle.dump((face_data,face_label), f)
f.close()

分類模型

程序參考了官方示例：https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py
一共有40個類，每個類10個樣本，共400個樣本。其中320個樣本用於訓練，40個用於驗證，剩下40個測試
注意給第一層指定input_shape，如果是MLP，代碼為：

model = Sequential()
# Dense(64) is a fully-connected layer with 64 hidden units.# in the first layer, you must specify the expected input data shape:# here, 20-dimensional vectors.
model.add(Dense(64, input_dim=20, init='uniform'))

后面可以不指定Dense的input shape

from __future__ import print_function
import numpy as np
import cPickle

np.random.seed(1337) # for reproducibililty

from keras.datasets import mnist
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from keras.utils import np_utils

# split data into train,vavlid and test
# train:320
# valid:40
# test:40
def split_data(fname):
    f = open(fname, 'rb')
    face_data,face_label = cPickle.load(f)

    X_train = np.empty((320, img_rows * img_cols))
    Y_train = np.empty(320, dtype=int)

    X_valid = np.empty((40, img_rows* img_cols))
    Y_valid = np.empty(40, dtype=int)

    X_test = np.empty((40, img_rows* img_cols))
    Y_test = np.empty(40, dtype=int)

    for i in range(40):
        X_train[i*8:(i+1)*8,:] = face_data[i*10:i*10+8,:]
        Y_train[i*8:(i+1)*8] = face_label[i*10:i*10+8]

        X_valid[i] = face_data[i*10+8,:]
        Y_valid[i] = face_label[i*10+8]

        X_test[i] = face_data[i*10+9,:]
        Y_test[i] = face_label[i*10+9]
    
    return (X_train, Y_train, X_valid, Y_valid, X_test, Y_test)

if __name__=='__main__':
    batch_size = 10
    nb_classes = 40
    nb_epoch = 12

    # input image dimensions
    img_rows, img_cols = 57, 47
    # number of convolutional filters to use
    nb_filters = 32
    # size of pooling area for max pooling
    nb_pool = 2
    # convolution kernel size
    nb_conv = 3

    (X_train, Y_train, X_valid, Y_valid, X_test, Y_test) = split_data('G:\data\olivettifaces.pkl')
    X_train = X_train.reshape(X_train.shape[0], 1, img_rows, img_cols)
    X_test = X_test.reshape(X_test.shape[0], 1, img_rows, img_cols)
    
    print('X_train shape:', X_train.shape)
    print(X_train.shape[0], 'train samples')
    print(X_test.shape[0], 'test samples')
    # convert label to binary class matrix
    Y_train = np_utils.to_categorical(Y_train, nb_classes)
    Y_test = np_utils.to_categorical(Y_test, nb_classes)

    model = Sequential()
    # 32 convolution filters , the size of convolution kernel is 3 * 3
    # border_mode can be 'valid' or 'full'
    #‘valid’only apply filter to complete patches of the image. 
    # 'full'  zero-pads image to multiple of filter shape to generate output of shape: image_shape + filter_shape - 1
    # when used as the first layer, you should specify the shape of inputs 
    # the first number means the channel of an input image, 1 stands for grayscale imgs, 3 for RGB imgs
    model.add(Convolution2D(nb_filters, nb_conv, nb_conv,
                            border_mode='valid',
                            input_shape=(1, img_rows, img_cols)))
    # use rectifier linear units : max(0.0, x)
    model.add(Activation('relu'))
    # second convolution layer with 32 filters of size 3*3
    model.add(Convolution2D(nb_filters, nb_conv, nb_conv))
    model.add(Activation('relu'))
    # max pooling layer, pool size is 2 * 2
    model.add(MaxPooling2D(pool_size=(nb_pool, nb_pool)))
    # drop out of max-pooling layer , drop out rate is 0.25 
    model.add(Dropout(0.25))
    # flatten inputs from 2d to 1d
    model.add(Flatten())
    # add fully connected layer with 128 hidden units
    model.add(Dense(128))
    model.add(Activation('relu'))
    model.add(Dropout(0.5))
    # output layer with softmax 
    model.add(Dense(nb_classes))
    model.add(Activation('softmax'))
    # use cross-entropy cost and adadelta to optimize params
    model.compile(loss='categorical_crossentropy', optimizer='adadelta')
    # train model with bath_size =10, epoch=12
    # set verbose=1 to show train info
    model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=nb_epoch,
          show_accuracy=True, verbose=1, validation_data=(X_test, Y_test))
    # evaluate on test set
    score = model.evaluate(X_test, Y_test, show_accuracy=True, verbose=0)
    print('Test score:', score[0])
    print('Test accuracy:', score[1])

結果：
准確率有97%