本文通过TensorFlow构建深度神经网络对MNIST手写体数字进行识别。学习其他内容参考TensorFlow 学习目录。
目录
一、MNIST数据集
链接：https://pan.baidu.com/s/1pX272g5fKo6zzc0SIuLB9w
提取码：kj9m
该数据集有60000个训练集，10000个测试集，数字有0~9十个数字。图片格式是灰度图，28*28大小，pkl中存储的形式是50000*784，10000*784，以及对应的标签。
<1>通过如下代码对数据集进行加载
# .py file: mnist_loader.py
import pickle
import gzip
import numpy as np
def load_data():
f = gzip.open('D:\dataset\mnist.pkl.gz', 'rb')
training_data, validation_data, test_data = pickle.load(f, encoding='bytes')
# print (training_data)
# print (validation_data)
# print (test_data)
f.close()
return (training_data, validation_data, test_data)
def vectorized_result(j):
e = np.zeros((10, 1))
e[j] = 1.0
return e
def load_data_wrapper():
tr_d, va_d, te_d = load_data()
#training_data
training_inputs = [np.reshape(x, (784, 1)) for x in tr_d[0]]
training_results = [vectorized_result(y) for y in tr_d[1]]
training_data = zip(training_inputs, training_results)
#validation_data
validation_inputs = [np.reshape(x, (784, 1)) for x in va_d[0]]
validation_results = [vectorized_result(y) for y in va_d[1]]
validation_data = zip(validation_inputs, validation_results)
#test_data
test_inputs = [np.reshape(x, (784, 1)) for x in te_d[0]]
test_data = zip(test_inputs, te_d[1])
return (training_data, validation_data, test_data)<2>通过如下代码对数据集进行划分，将validation的10000个样例划分到训练集中，这样训练集就有60000张图片。
# .py file: get_Dataset.py
import numpy as np
import mnist_loader
def get_Dataset(name='mnist'):
if name == 'mnist':
t, v, tt = mnist_loader.load_data_wrapper()
validation_data = list(v)
training_data = list(t) + validation_data
testing_data = list(tt)
len_t = len(training_data)
len_tdi = len(training_data[0][0])
len_tl = len(training_data[0][1])
x_train = np.zeros((len_t, len_tdi))
y_train = np.zeros((len_t, len_tl))
for i in range(len_t):
x_train[i] = np.array(training_data[i][0]).transpose()
y_train[i] = np.array(training_data[i][1]).transpose()
len_tt = len(testing_data)
x_test = np.zeros((len_tt, len_tdi))
y_test = np.zeros(len_tt)
for i in range(len_tt):
x_test[i] = np.array(testing_data[i][0]).transpose()
y_test[i] = testing_data[i][1]
return x_train, y_train, x_test, y_test二、构建深度神经网络
<1>入口 placeholder
inputs = tf.placeholder(dtype=tf.float32, shape=[None, 784], name='inputs')
labels = tf.placeholder(dtype=tf.float32, shape=[None, 10], name='labels')
predict_labels = tf.placeholder(dtype=tf.int64, shape=[None], name='predict_labels')其中inputs表示输入的784维度的灰度图的向量拉伸，labels是一个one-hot向量，predict_labels不是one-hot向量，是直接的0~9数字。
<2>卷积层定义
def conv2d(x, id, filter_size):
with tf.variable_scope('conv2d_'+str(id)):
w = tf.get_variable(name='filter', initializer=tf.truncated_normal(shape=filter_size, stddev=0.1), dtype=tf.float32)
b = tf.get_variable(name='bias', initializer=tf.constant(0.1, shape=[filter_size[3]], dtype=tf.float32))
z = tf.nn.relu(tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME') + b)
return z这里用到了tf.variable_scope()+tf.get_variable()的方式，具体参考TensorFlow 共享变量。
<3>池化层定义，此处采用最大值池化
def pooling(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')<4>Tensor拉伸操作，flatten操作：将多通道的tensor转化为一维向量
def flatten(x):
l = x.get_shape()[1:]
length = l[1]*l[2]*l[0]
x = tf.reshape(x, shape=[-1, length])
return x<5>全连接层
def fully(x, id, neurons):
with tf.variable_scope('fully'+str(id)):
L = int(x.get_shape()[1])
w = tf.get_variable(name='w', initializer=tf.truncated_normal(shape=[L, neurons], stddev=0.1), dtype=tf.float32)
b = tf.get_variable(name='b', initializer=tf.constant(0.1, shape=[neurons], dtype=tf.float32))
z = tf.nn.relu(tf.nn.xw_plus_b(x, w, b))
return z三、网络结构的设计
- 图像输入[batchsize, 784]
- net为图像reshape之后的结果，[batchsize, 28, 28, 1]，因为是灰度图，所以通道数为 1。
- 建立一个大小为5*5，进入通道数为 1，输出通道数为64的卷积核对net进行卷积操作
- 对net进行一次2*2大小的池化操作
- 建立一个大小为3*3，计入通道数为 64，输出通道数为64的卷积核对net进行卷积操作
- 对net进行一次2*2大小的池化操作
- 将net进行拉伸，变为一维向量（加上batchsize其实是二维）
- 加入一层神经元个数为512的全连接层
- 加入一层神经元个数为10的全连接层作为输出层
net = tf.reshape(inputs, shape=[-1, 28, 28, 1])
net = conv2d(net, 1, [5, 5, 1, 64])
net = pooling(net)
net = conv2d(net, 2, [3, 3, 64, 64])
net = pooling(net)
net = flatten(net)
net = fully(net, 3, 512)
logits = fully(net, 4, 10)四、定义损失函数，以及优化方式
采用交叉熵作为损失函数，采用AdamOptimizer作为优化方式。
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)五、预测
correct_prediction = tf.equal(tf.argmax(logits, 1), predict_labels)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))六、定义Epochs、batch_size进行训练
epochs = 100
batch_size = 32
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
for batch in range(int(x_train.shape[0]/batch_size)):
batch_xs = x_train[batch*batch_size: (batch+1)*batch_size]
batch_ys = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batch_xs,
labels: batch_ys
}
sess.run(train_step, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, predict_labels: y_test})
print ("Epoch ", epoch+1, " acc=", acc*100, "%")输出结果
Epoch 1 acc= 28.949999809265137 %
Epoch 2 acc= 57.270002365112305 %
Epoch 3 acc= 57.74000287055969 %
Epoch 4 acc= 57.88000226020813 %
Epoch 5 acc= 67.66999959945679 %
Epoch 6 acc= 77.30000019073486 %
Epoch 7 acc= 87.48999834060669 %
Epoch 8 acc= 87.58999705314636 %
Epoch 9 acc= 87.6200020313263 %
Epoch 10 acc= 87.61000037193298 %
Epoch 11 acc= 87.58000135421753 %
Epoch 12 acc= 87.80999779701233 %
Epoch 13 acc= 87.77999877929688 %
Epoch 14 acc= 87.87000179290771 %
Epoch 15 acc= 87.84000277519226 %
Epoch 16 acc= 87.8000020980835 %
Epoch 17 acc= 87.73999810218811 %
Epoch 18 acc= 87.76999711990356 %
Epoch 19 acc= 87.77999877929688 %
Epoch 20 acc= 87.87000179290771 %
Epoch 21 acc= 87.84000277519226 %
Epoch 22 acc= 87.80999779701233 %
Epoch 23 acc= 87.69000172615051 %
Epoch 24 acc= 87.74999976158142 %
Epoch 25 acc= 87.80999779701233 %
Epoch 26 acc= 98.86999726295471 %
Epoch 27 acc= 99.16999936103821 %
Epoch 28 acc= 99.19000267982483 %
Epoch 29 acc= 99.1599977016449 %
Epoch 30 acc= 99.15000200271606 %
Epoch 31 acc= 99.08000230789185 %
Epoch 32 acc= 99.14000034332275 %
Epoch 33 acc= 99.16999936103821 %
Epoch 34 acc= 99.1599977016449 %
Epoch 35 acc= 99.14000034332275 %
Epoch 36 acc= 99.04999732971191 %
Epoch 37 acc= 99.18000102043152 %
Epoch 38 acc= 99.22000169754028 %
Epoch 39 acc= 99.19000267982483 %
Epoch 40 acc= 99.19999837875366 %
Epoch 41 acc= 99.26999807357788 %
Epoch 42 acc= 99.11999702453613 %
Epoch 43 acc= 99.22000169754028 %
Epoch 44 acc= 99.19999837875366 %
Epoch 45 acc= 99.26999807357788 %
Epoch 46 acc= 99.26999807357788 %
Epoch 47 acc= 99.09999966621399 %
Epoch 48 acc= 99.1100013256073 %
Epoch 49 acc= 99.14000034332275 %
Epoch 50 acc= 99.01999831199646 %
Epoch 51 acc= 99.12999868392944 %
Epoch 52 acc= 98.96000027656555 %
Epoch 53 acc= 99.11999702453613 %
Epoch 54 acc= 99.11999702453613 %
Epoch 55 acc= 99.22999739646912 %
Epoch 56 acc= 99.26000237464905 %
Epoch 57 acc= 99.25000071525574 %
Epoch 58 acc= 99.23999905586243 %
Epoch 59 acc= 99.19000267982483 %
Epoch 60 acc= 99.21000003814697 %
Epoch 61 acc= 99.23999905586243 %
Epoch 62 acc= 99.25000071525574 %
Epoch 63 acc= 99.19000267982483 %
Epoch 64 acc= 99.16999936103821 %
Epoch 65 acc= 99.04000163078308 %
Epoch 66 acc= 99.12999868392944 %
Epoch 67 acc= 99.1100013256073 %
Epoch 68 acc= 99.22000169754028 %
Epoch 69 acc= 99.18000102043152 %
Epoch 70 acc= 99.22000169754028 %
Epoch 71 acc= 99.04000163078308 %
Epoch 72 acc= 99.22000169754028 %
Epoch 73 acc= 99.23999905586243 %
Epoch 74 acc= 99.26999807357788 %
Epoch 75 acc= 99.05999898910522 %
Epoch 76 acc= 99.19000267982483 %
Epoch 77 acc= 99.21000003814697 %
Epoch 78 acc= 99.19999837875366 %
Epoch 79 acc= 99.22000169754028 %
Epoch 80 acc= 99.23999905586243 %
Epoch 81 acc= 99.15000200271606 %
Epoch 82 acc= 99.25000071525574 %
Epoch 83 acc= 99.25000071525574 %
Epoch 84 acc= 99.1599977016449 %
Epoch 85 acc= 99.22999739646912 %
Epoch 86 acc= 99.21000003814697 %
Epoch 87 acc= 99.2900013923645 %
Epoch 88 acc= 99.21000003814697 %
Epoch 89 acc= 99.27999973297119 %
Epoch 90 acc= 99.18000102043152 %
Epoch 91 acc= 99.04000163078308 %
Epoch 92 acc= 99.26999807357788 %
Epoch 93 acc= 99.12999868392944 %
Epoch 94 acc= 99.2900013923645 %
Epoch 95 acc= 99.32000041007996 %
Epoch 96 acc= 99.25000071525574 %
Epoch 97 acc= 99.2900013923645 %
Epoch 98 acc= 99.26000237464905 %
Epoch 99 acc= 99.16999936103821 %
Epoch 100 acc= 99.27999973297119 %
七、附件：完整代码
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
def conv2d(x, id, filter_size):
with tf.variable_scope('conv2d_'+str(id)):
w = tf.get_variable(name='filter', initializer=tf.truncated_normal(shape=filter_size, stddev=0.1), dtype=tf.float32)
b = tf.get_variable(name='bias', initializer=tf.constant(0.1, shape=[filter_size[3]], dtype=tf.float32))
z = tf.nn.relu(tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME') + b)
return z
def pooling(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
def flatten(x):
l = x.get_shape()[1:]
length = l[1]*l[2]*l[0]
x = tf.reshape(x, shape=[-1, length])
return x
def fully(x, id, neurons):
with tf.variable_scope('fully'+str(id)):
L = int(x.get_shape()[1])
w = tf.get_variable(name='w', initializer=tf.truncated_normal(shape=[L, neurons], stddev=0.1), dtype=tf.float32)
b = tf.get_variable(name='b', initializer=tf.constant(0.1, shape=[neurons], dtype=tf.float32))
z = tf.nn.relu(tf.nn.xw_plus_b(x, w, b))
return z
inputs = tf.placeholder(dtype=tf.float32, shape=[None, 784], name='inputs')
labels = tf.placeholder(dtype=tf.float32, shape=[None, 10], name='labels')
predict_labels = tf.placeholder(dtype=tf.int64, shape=[None], name='predict_labels')
net = tf.reshape(inputs, shape=[-1, 28, 28, 1])
net = conv2d(net, 1, [5, 5, 1, 64])
net = pooling(net)
net = conv2d(net, 2, [3, 3, 64, 64])
net = pooling(net)
net = flatten(net)
net = fully(net, 3, 512)
logits = fully(net, 4, 10)
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)
correct_prediction = tf.equal(tf.argmax(logits, 1), predict_labels)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
epochs = 100
batch_size = 32
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(epochs):
for batch in range(int(x_train.shape[0]/batch_size)):
batch_xs = x_train[batch*batch_size: (batch+1)*batch_size]
batch_ys = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batch_xs,
labels: batch_ys
}
sess.run(train_step, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, predict_labels: y_test})
print ("Epoch ", epoch+1, " acc=", acc*100, "%")