TensorFlow【极简】CNN

版权声明：本文为博主原创文章，转载请注明出处。 https://blog.csdn.net/Yellow_python/article/details/86563960

单层卷积神经网络

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

"""加载样本集：手写数字"""
mnist = input_data.read_data_sets('data/', one_hot=True)

"""卷积神经网络"""
# 输入层
X = tf.placeholder('float', [None, 28, 28, 1])  # 批量、高、宽、通道数
y = tf.placeholder('float', [None, 10])  # 10分类（0~9）

# 卷积层
W_conv = tf.Variable(tf.truncated_normal([5, 5, 1, 32], stddev=0.1))  # 高、宽、通道数、filter个数
b_conv = tf.Variable(tf.constant(.1, shape=[32]))  # filter个数
h_conv = tf.nn.conv2d(input=X, filter=W_conv, strides=[1, 1, 1, 1], padding='SAME') + b_conv  # 滑窗步幅：水平1、垂直1

# ReLU激活
h_relu = tf.nn.relu(h_conv)

# 池化层
h_pool = tf.nn.max_pool(h_relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')  # 滑窗大小：水平2、垂直2；滑窗步幅：水平2、垂直2

# 平化层
h_flat = tf.reshape(h_pool, [-1, 14 * 14 * 32])

# 全连接层1
W_fc1 = tf.Variable(tf.truncated_normal([14 * 14 * 32, 256], stddev=.1))  # 指定256个神经元
b_fc1 = tf.Variable(tf.constant(.1, shape=[256]))
h_fc1 = tf.matmul(h_flat, W_fc1) + b_fc1
h_fc1 = tf.nn.relu(h_fc1)

# 全连接层2
W_fc2 = tf.Variable(tf.truncated_normal([256, 10], stddev=.1))
b_fc2 = tf.Variable(tf.constant(.1, shape=[10]))
h_fc2 = tf.matmul(h_fc1, W_fc2) + b_fc2

# softmax交叉熵损失
loss = tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=h_fc2)

# Adam优化器
optimizer = tf.train.AdamOptimizer().minimize(loss)

"""精度"""
prediction = tf.equal(tf.argmax(h_fc2, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(prediction, 'float'))

"""运行"""
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
batch_size = 50
for i in range(501):
    batch = mnist.train.next_batch(batch_size)  # 分批
    inputs = batch[0].reshape([batch_size, 28, 28, 1])
    labels = batch[1]
    optimizer.run(session=sess, feed_dict={X: inputs, y: labels})  # 训练
    if i % 10 == 0:
        train_accuracy = accuracy.eval(session=sess, feed_dict={X: inputs, y: labels})
        print('step %d, accuracy %g' % (i, train_accuracy))

知识补充

1、图像特征

from skimage import data  # conda install pillow
import numpy as np, matplotlib.pyplot as mp
cat = data.chelsea()
print(cat.shape)  # 高、宽、通道：(300, 451, 3)
# RGB通道
cat_R = cat.copy()  # 红色通道
cat_R[:, :, 1] = cat_R[:, :, 2] = 0  # 绿、蓝通道像素设0
cat_G = cat.copy()  # 绿色通道
cat_G[:, :, 0] = cat_G[:, :, 2] = 0  # 红、蓝通道像素设0
cat_B = cat.copy()  # 蓝色通道
cat_B[:, :, 0] = cat_B[:, :, 1] = 0  # 红、绿通道像素设0
# 可视化
plot_image = np.concatenate((cat, cat_R, cat_G, cat_B), axis=1)
mp.imshow(plot_image)
mp.show()

2、mnist数据集

from tensorflow.examples.tutorials.mnist import input_data
import numpy as np, matplotlib.pyplot as mp
mnist = input_data.read_data_sets('data/', one_hot=True)
train_images = mnist.train.images  # shape(55000, 784)
train_labels = mnist.train.labels  # shape(55000, 10)
test_images = mnist.test.images  # shape(10000, 784)
testl_labels = mnist.test.labels  # shape(10000, 10)
# 可视化
for i in range(7):
    image = np.reshape(train_images[i, :], (28, 28))
    label = np.argmax(train_labels[i, :])
    mp.subplot(1, 7, i + 1)
    mp.axis('off')
    mp.imshow(image, cmap=mp.get_cmap('gray'))
    mp.title(label)
mp.show()

3、TensorFlow语法

• 所有变量必须经过初始化才能启用

import tensorflow as tf
"""创建tf格式的变量"""
a = tf.Variable([[1], [2]])  # <tf.Variable 'Variable:0' shape=(2, 1) dtype=int32_ref>
b = tf.Variable([[1, 2, 3]])  # <tf.Variable 'Variable_1:0' shape=(1, 3) dtype=int32_ref>
c = tf.matmul(a, b)  # Tensor("MatMul:0", shape=(2, 3), dtype=int32)
print(a, b, c, sep='\n')
"""全局变量初始化"""
with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    print(sess.run(c))

4、模型评估

• 完整CNN示例：2层卷积、再加上dropout

import tensorflow as tf, matplotlib.pyplot as mp
from tensorflow.examples.tutorials.mnist import input_data

"""加载样本集：手写数字"""
mnist = input_data.read_data_sets('data/', one_hot=True)

"""卷积神经网络"""
# 卷积、ReLU、池化
def conv(X, channels, filters):
    W_conv = tf.Variable(tf.truncated_normal([5, 5, channels, filters], stddev=0.1))
    b_conv = tf.Variable(tf.constant(.1, shape=[filters]))
    h_conv = tf.nn.conv2d(input=X, filter=W_conv, strides=[1, 1, 1, 1], padding='SAME') + b_conv
    h_relu = tf.nn.relu(h_conv)
    return tf.nn.max_pool(h_relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# 输入层
X = tf.placeholder('float', [None, 28, 28, 1])
y = tf.placeholder('float', [None, 10])
# 2次卷积
h_pool1 = conv(X, 1, 32)
h_pool2 = conv(h_pool1, 32, 64)
# 平化层
h_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
# 全连接层1
W_fc1 = tf.Variable(tf.truncated_normal([7 * 7 * 64, 512], stddev=.1))
b_fc1 = tf.Variable(tf.constant(.1, shape=[512]))
h_fc1 = tf.nn.relu(tf.matmul(h_flat, W_fc1) + b_fc1)
# Dropout减少过拟合
keep_prob = tf.placeholder('float')
h_drop = tf.nn.dropout(h_fc1, keep_prob)
# 全连接层2
W_fc2 = tf.Variable(tf.truncated_normal([512, 10], stddev=.1))
b_fc2 = tf.Variable(tf.constant(.1, shape=[10]))
h_fc2 = tf.matmul(h_drop, W_fc2) + b_fc2
# softmax交叉熵损失
h_softmax = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=h_fc2))
# Adam优化器
optimizer = tf.train.AdamOptimizer().minimize(h_softmax)
# 精度
prediction = tf.equal(tf.argmax(h_fc2, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(prediction, 'float'))

"""运行和预测"""
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
batch_size = 50
for i in range(101):
    batch = mnist.train.next_batch(batch_size)
    inputs = batch[0].reshape([batch_size, 28, 28, 1])
    labels = batch[1]
    feed_dict = {X: inputs, y: labels, keep_prob: .9}
    optimizer.run(feed_dict, sess)
    if i % 10 == 0:
        loss = h_softmax.eval(feed_dict, sess)
        accu = accuracy.eval(feed_dict, sess)
        print('step %3d; loss %.2f; accuracy %g' % (i, loss, accu))
    if i == 100:
        real = tf.argmax(y, 1)
        realities = tf.argmax(y, 1).eval(feed_dict, sess)
        predictions = tf.argmax(h_fc2, 1).eval(feed_dict, sess)
        for i in range(8):
            mp.subplot(1, 8, i + 1)
            mp.axis('off')
            mp.imshow(inputs[i].reshape(28, 28), cmap=mp.get_cmap('gray'))
            mp.title('real %d; pred %d' % (realities[i], predictions[i]))
        mp.show()

step 0; loss 6.68; accuracy 0.22
step 10; loss 2.36; accuracy 0.32
step 20; loss 1.52; accuracy 0.58
step 30; loss 0.93; accuracy 0.76
step 40; loss 0.54; accuracy 0.8
step 50; loss 0.43; accuracy 0.88
step 60; loss 0.20; accuracy 0.98
step 70; loss 0.47; accuracy 0.86
step 80; loss 0.39; accuracy 0.92
step 90; loss 0.28; accuracy 0.92
step 100; loss 0.21; accuracy 0.96

5、注释

en	cn
convolution	卷积
placeholder	占位符
constant	常量
stride	步幅
optimize	优化
truncate	截短