最近在抽时间学习TensorFlow这个DL库的使用，学的断断续续的，看官网上第一个案例就是训练手写字符识别。我之前在做Weibo.cn验证码识别的时候，自己搞了一个数据集，当时用的c++库tiny-dnn进行训练的(见：验证码破解技术四部曲之使用卷积神经网络（四）)，现在我把它移植到TensorFlow上试试。
完整代码见：weibo.cn/tensorflow-impl
使用的库
TensorFlow-1.0
scikit-learn-0.18
pillow
加载数据集
数据集下载地址：training_set.zip
解压过后如下图： 
我把同一类的图片放到了一个文件夹里，文件夹的名字也就是图片的label，打开文件夹后可以看到字符的图片信息。 
下面，我们把数据加载到一个pickle文件里面，它需要有train_dataset、train_labels、test_dataset、test_labels四个变量代表训练集和测试集的数据和标签。
此外，还需要有个label_map，用来把训练的标签和实际的标签对应，比如说3对应字母M，4对应字母N。
此部分的代码见：load_models.py。注：很多的代码参考自udacity的deeplearning课程。
首先根据文件夹的来加载所有的数据，index代表训练里的标签，label代表实际的标签，使用PIL读取图片，并转换成numpy数组。
import numpy as np
import os
from PIL import Image
def load_dataset():
dataset = []
labelset = []
label_map = {}
base_dir = "../trainer/training_set/" # 数据集的位置
labels = os.listdir(base_dir)
for index, label in enumerate(labels):
if label == "ERROR" or label == ".DS_Store":
continue
print "loading:", label, "index:", index
try:
image_files = os.listdir(base_dir + label)
for image_file in image_files:
image_path = base_dir + label + "/" + image_file
im = Image.open(image_path).convert('L')
dataset.append(np.asarray(im, dtype=np.float32))
labelset.append(index)
label_map[index] = label
except: pass
return np.array(dataset), np.array(labelset), label_map
dataset, labelset, label_map = load_dataset()接下来，把数据打乱。
def randomize(dataset, labels):
permutation = np.random.permutation(labels.shape[0])
shuffled_dataset = dataset[permutation, :, :]
shuffled_labels = labels[permutation]
return shuffled_dataset, shuffled_labels
dataset, labelset = randomize(dataset, labelset)然后使用scikit-learn的函数，把训练集和测试集分开。
from sklearn.model_selection import train_test_split
train_dataset, test_dataset, train_labels, test_labels = train_test_split(dataset, labelset)在TensorFlow官网给的例子中，会把label进行One-Hot Encoding，并把28*28的图片转换成了一维向量(784)。如下图，查看官网例子的模型。
我也把数据转换了一下，把32*32的图片转换成一维向量(1024)，并对标签进行One-Hot Encoding。
def reformat(dataset, labels, image_size, num_labels):
dataset = dataset.reshape((-1, image_size * image_size)).astype(np.float32)
# Map 1 to [0.0, 1.0, 0.0 ...], 2 to [0.0, 0.0, 1.0 ...]
labels = (np.arange(num_labels) == labels[:, None]).astype(np.float32)
return dataset, labels
train_dataset, train_labels = reformat(train_dataset, train_labels, 32, len(label_map))
test_dataset, test_labels = reformat(test_dataset, test_labels, 32, len(label_map))
print "train_dataset:", train_dataset.shape
print "train_labels:", train_labels.shape
print "test_dataset:", test_dataset.shape
print "test_labels:", test_labels.shape转换后，格式就和minist一样了。
最后，把数据保存到save.pickle里面。
save = {
'train_dataset': train_dataset,
'train_labels': train_labels,
'test_dataset': test_dataset,
'test_labels': test_labels,
'label_map': label_map
}
with open("save.pickle", 'wb') as f:
pickle.dump(save, f)验证数据集加载是否正确
加载完数据后，需要验证一下数据是否正确。我选择的方法很简单，就是把trainset的第1个(或者第2个、第n个)图片打开，看看它的标签和看到的能不能对上。
import cPickle as pickle
from PIL import Image
import numpy as np
def check_dataset(dataset, labels, label_map, index):
data = np.uint8(dataset[index]).reshape((32, 32))
i = np.argwhere(labels[index] == 1)[0][0]
im = Image.fromarray(data)
im.show()
print "label:", label_map[i]
if __name__ == '__main__':
with open("save.pickle", 'rb') as f:
save = pickle.load(f)
train_dataset = save['train_dataset']
train_labels = save['train_labels']
test_dataset = save['test_dataset']
test_labels = save['test_labels']
label_map = save['label_map']
# check if the image is corresponding to it's label
check_dataset(train_dataset, train_labels, label_map, 0)运行后，可以看到第一张图片是Y，标签也是正确的。
训练
数据加载好了之后，就可以开始训练了，训练的网络就使用TensorFlow官网在Deep MNIST for Experts里提供的就好了。
此部分的代码见：train.py。
先加载一下模型:
import cPickle as pickle
import numpy as np
import tensorflow as tf
with open("save.pickle", 'rb') as f:
save = pickle.load(f)
train_dataset = save['train_dataset']
train_labels = save['train_labels']
test_dataset = save['test_dataset']
test_labels = save['test_labels']
label_map = save['label_map']
image_size = 32
num_labels = len(label_map)
print "train_dataset:", train_dataset.shape
print "train_labels:", train_labels.shape
print "test_dataset:", test_dataset.shape
print "test_labels:", test_labels.shape
print "num_labels:", num_labelsminist的数据都是28*28的，把里面的网络改完了之后，如下：
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
graph = tf.Graph()
with graph.as_default():
x = tf.placeholder(tf.float32, shape=[None, image_size * image_size])
y_ = tf.placeholder(tf.float32, shape=[None, num_labels])
x_image = tf.reshape(x, [-1, 32, 32, 1])
# First Convolutional Layer
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
# Second Convolutional Layer
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
# Densely Connected Layer
W_fc1 = weight_variable([image_size / 4 * image_size / 4 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, image_size / 4 * image_size / 4 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
# Dropout
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# Readout Layer
W_fc2 = weight_variable([1024, num_labels])
b_fc2 = bias_variable([num_labels])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))主要改动就是输入层把28*28改成了image_size*image_size（32*32），然后第三层的全连接网络把7*7改成了image_size/4*image_size/4（8*8），以及把10（手写字符一共10类）改成了num_labels。
然后训练，我这里把batch_size改成了128，训练批次改少了。
batch_size = 128
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print("Initialized")
for step in range(2001):
offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
# Generate a minibatch.
batch_data = train_dataset[offset:(offset + batch_size), :]
batch_labels = train_labels[offset:(offset + batch_size), :]
if step % 50 == 0:
train_accuracy = accuracy.eval(feed_dict={
x: batch_data, y_: batch_labels, keep_prob: 1.0})
test_accuracy = accuracy.eval(feed_dict={
x: test_dataset, y_: test_labels, keep_prob: 1.0})
print("Step %d, Training accuracy: %g, Test accuracy: %g" % (step, train_accuracy, test_accuracy))
train_step.run(feed_dict={x: batch_data, y_: batch_labels, keep_prob: 0.5})
print("Test accuracy: %g" % accuracy.eval(feed_dict={
x: test_dataset, y_: test_labels, keep_prob: 1.0}))运行，可以看到识别率在不断的上升。
最后，有了接近98%的识别率，只有4000个训练数据，感觉不错了。
