Tensorflow是一个非常好用的deep learning框架
学完了cs231n，大概就可以写一个CNN做一下MNIST了
tensorflow具体原理可以参见它的官方文档
然后CNN的原理可以直接学习cs231n的课程。
另外这份代码本地跑得奇慢。。估计用gpu会快很多。
import loaddata import tensorflow as tf #生成指定大小符合标准差为0.1的正态分布的矩阵 def weight_variable(shape): initial = tf.truncated_normal(shape, stddev = 0.1) return tf.Variable(initial) #生成偏移变量 def bias_variable(shape): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial) #做W与x的卷积运算，跨度为1，zero-padding补全边界（使得最后结果大小一致） def conv2d(x, W): return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') #做2x2的max池化运算，使结果缩小4倍（面积上） def max_pool_2x2(x): return tf.nn.max_pool(x, ksize = [1, 2, 2, 1], strides=[1, 2, 2, 1], padding = 'SAME') #导入数据 mnist = loaddata.read_data_sets('MNIST_data', one_hot=True) x = tf.placeholder("float", shape=[None, 784]) y_ = tf.placeholder("float", shape=[None, 10]) #filter取5x5的范围，因为mnist为单色，所以第三维是1，卷积层的深度为32 W_conv1 = weight_variable([5, 5, 1, 32]) b_conv1 = bias_variable([32]) #将输入图像变成28*28*1的形式，来进行卷积 x_image = tf.reshape(x, [-1, 28, 28, 1]) #卷积运算，activation为relu h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) #池化运算 h_pool1 = max_pool_2x2(h_conv1) #第二个卷积层，深度为64，filter仍然取5x5 W_conv2 = weight_variable([5, 5, 32, 64]) b_conv2 = bias_variable([64]) #做同样的运算 h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) h_pool2 = max_pool_2x2(h_conv2) #full-connected层，将7*7*64个神经元fc到1024个神经元上去 W_fc1 = weight_variable([7*7*64, 1024]) b_fc1 = bias_variable([1024]) #将h_pool2（池化后的结果）打平后，进行fc运算 h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64]) h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) #防止过拟合，fc层进行dropout处理，参数为0.5 keep_prob = tf.placeholder("float") h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) #第二个fc层，将1024个神经元fc到10个最终结果上去（分别对应0~9） W_fc2 = weight_variable([1024, 10]) b_fc2 = bias_variable([10]) #最后结果 y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) #误差函数使用交叉熵 cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv)) #梯度下降使用adam算法 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) #正确率处理 correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) #初始化 sess = tf.Session() sess.run(tf.initialize_all_variables()) #进行训练 for i in range(20000): batch = mnist.train.next_batch(50) if i%100 == 0: train_accuracy = sess.run(accuracy, feed_dict = { x:batch[0], y_:batch[1], keep_prob : 1.0}) print("step %d, accuracy %g" % (i, train_accuracy)) sess.run(train_step, feed_dict={x:batch[0], y_:batch[1], keep_prob:0.5}) #输出最终结果 print(sess.run(accuracy, feed_dict={ x:mnist.test.images, y_:mnist.test.labels, keep_prob:1.0}))