Tensorflow + AlexNet + MNIST 填坑小记
最近，帮老婆debug过程中，也学习了一些tensorflow的相关知识。其中，花了比较多的实验研究如何使用tf.dataset来导入数据。以下是对一个demo的小记录，完成的工作是在tensorflow上实现AlexNet，使用的数据为MNIST。
由于国内网络原因，直接用tf中的函数读取mnist数据集较慢，我选择使用keras.datasets中import mnist数据集。这就导致后续的训练中，无法直接使用mnist.train.next_batch(batch_size)来生成mnist的训练数据。因此，选择使用tf.dataset将读取的numpy数据转化为tf所需格式。
读取数据代码如下：
from keras.datssets import mnist
from keras.utils import to_categorical
def load_mnist(image_size):
(x_train,y_train),(x_test,y_test) = mnist.load_data()
train_image = [cv2.cvtColor(cv2.resize(img,(image_size,image_size)),cv2.COLOR_GRAY2BGR) for img in x_train]
test_image = [cv2.cvtColor(cv2.resize(img,(image_size,image_size)),cv2.COLOR_GRAY2BGR) for img in x_test]
train_image = np.asarray(train_image, 'f')
test_image = np.asarray(test_image, 'f')
train_label = to_categorical(y_train)
test_label = to_categorical(y_test)
print('finish loading data!')
return train_image, train_label, test_image, test_labelload_mnist函数完成了mnist数据的读取，并将单通道转化为3通道（使用opencv函数），将label转化为onehot编码。
接下来，进行网络的初始化。网络结构转载自老婆的博客。
首先定义几个层结构：
# define layers
def conv(x, kernel, strides, b):
return tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, kernel, strides, padding = 'SAME'), b))
def max_pooling(x, kernel, strides):
return tf.nn.max_pool(x, kernel, strides, padding = 'VALID')
def fc(x, w, b):
return tf.nn.relu(tf.add(tf.matmul(x,w),b))其次，定义网络初始权值：
# define variables
weights = {
'wc1':tf.Variable(tf.random_normal([11,11,3,96], dtype=tf.float32, stddev=0.1), name='weights1'),
'wc2':tf.Variable(tf.random_normal([5,5,96,256], dtype=tf.float32, stddev=0.1), name='weights2'),
'wc3':tf.Variable(tf.random_normal([3,3,256,384], dtype=tf.float32, stddev=0.1), name='weights3'),
'wc4':tf.Variable(tf.random_normal([3,3,384,384], dtype=tf.float32, stddev=0.1), name='weights4'),
'wc5':tf.Variable(tf.random_normal([3,3,384,256], dtype=tf.float32, stddev=0.1), name='weights5'),
'wd1':tf.Variable(tf.random_normal([6*6*256, 4096], dtype=tf.float32, stddev=0.1), name='weights_fc1'),
'wd2':tf.Variable(tf.random_normal([4096, 1000], dtype=tf.float32, stddev=0.1), name='weights_fc2'),
'wd3':tf.Variable(tf.random_normal([1000, n_output], dtype=tf.float32, stddev=0.1), name='weights_fc3'),
}
bias = {
'bc1':tf.Variable(tf.zeros([96]), name='bias1'),
'bc2':tf.Variable(tf.zeros([256]), name='bias2'),
'bc3':tf.Variable(tf.zeros([384]), name='bias3'),
'bc4':tf.Variable(tf.zeros([384]), name='bias4'),
'bc5':tf.Variable(tf.zeros([256]), name='bias5'),
'bd1':tf.Variable(tf.zeros([4096]), name='bias_fc1'),
'bd2':tf.Variable(tf.zeros([1000]), name='bias_fc2'),
'bd3':tf.Variable(tf.zeros([n_output]), name='bias_fc3'),
}
strides = {
'sc1':[1,4,4,1],
'sc2':[1,1,1,1],
'sc3':[1,1,1,1],
'sc4':[1,1,1,1],
'sc5':[1,1,1,1],
'sp1':[1,2,2,1],
'sp2':[1,2,2,1],
'sp3':[1,2,2,1]
}
pooling_size = {
'kp1':[1,3,3,1],
'kp2':[1,3,3,1],
'kp3':[1,3,3,1]
}随后搭建网络：
#build model
def alexnet(inputs, weights, bias, strides, pooling_size, keep_prob):
with tf.name_scope('conv1'):
conv1 = conv(inputs, weights['wc1'], strides['sc1'], bias['bc1'])
print_layer(conv1)
with tf.name_scope('pool1'):
pool1 = max_pooling(conv1, pooling_size['kp1'], strides['sp1'])
print_layer(pool1)
with tf.name_scope('conv2'):
conv2 = conv(pool1, weights['wc2'], strides['sc2'], bias['bc2'])
print_layer(conv2)
with tf.name_scope('pool2'):
pool2 = max_pooling(conv2, pooling_size['kp2'], strides['sp2'])
print_layer(pool2)
with tf.name_scope('conv3'):
conv3 = conv(pool2, weights['wc3'], strides['sc3'], bias['bc3'])
print_layer(conv3)
with tf.name_scope('conv4'):
conv4 = conv(conv3, weights['wc4'], strides['sc4'], bias['bc4'])
print_layer(conv4)
with tf.name_scope('conv5'):
conv5 = conv(conv4, weights['wc5'], strides['sc5'], bias['bc5'])
print_layer(conv5)
with tf.name_scope('pool3'):
pool3 = max_pooling(conv5, pooling_size['kp3'], strides['sp3'])
print_layer(pool3)
flatten = tf.reshape(pool3, [-1,6*6*256])
with tf.name_scope('fc1'):
fc1 = fc(flatten, weights['wd1'], bias['bd1'])
fc1_drop = tf.nn.dropout(fc1, keep_prob)
print_layer(fc1_drop)
with tf.name_scope('fc2'):
fc2 = fc(fc1_drop, weights['wd2'], bias['bd2'])
fc2_drop = tf.nn.dropout(fc2, keep_prob)
print_layer(fc2_drop)
with tf.name_scope('fc3'):
outputs = tf.add(tf.matmul(fc2,weights['wd3']),bias['bd3'])
print_layer(outputs)
return outputs接下来将读取的numpy类型数据处理为tf.dataset的数据格式。由于数据量较大，超过tf规定的2GB要求，因此需要使用可初始话的迭代器来进行数据读取。读取方式如下：
x_dataset = tf.placeholder(train_image.dtype, train_image.shape)
y_dataset = tf.placeholder(train_label.dtype, train_label.shape)
train_dataset = tf.data.Dataset.from_tensor_slices((x_dataset,y_dataset)).repeat().batch(batch_size)
train_iterator = train_dataset.make_initializable_iterator()
features, labels = train_iterator.get_next()其中，使用tf.placeholder占位符来进行初始化，并使用tf.data.Dataset.from_tensor_slices设置dataset。为了使代码能够运行多个epoch，使用repeat函数创建多个周期的数据集（具体使用参照tensorflow官网）。
然后，建立目标函数，选定优化方式，并设定计算acc方法。
keep_prob = tf.placeholder(tf.float32)
pred = alexnet(features, weights, bias, strides, pooling_size, keep_prob)
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=labels))
train_step = tf.train.AdamOptimizer(learning_rate = lr).minimize(loss)
correct = tf.equal(tf.argmax(labels,1), tf.argmax(pred,1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))这样就可以开始训练了，训练的代码如下：
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(train_iterator.initializer, feed_dict={x_dataset: train_image, y_dataset: train_label})
for epoch in range(epochs):
_, loss_value, acc = sess.run([train_step, loss, accuracy], feed_dict = {keep_prob: dropout_rate})
if (epoch+1)%100 == 0:
print("Iter: {}, Loss: {:.4f}, Acc: {:.4f}".format(epoch, loss_value, acc))
print ("Optimization Finished!")经过10000次左右的训练，所得效果如下：
