diff --git a/homework7/.idea/.gitignore b/homework7/.idea/.gitignore new file mode 100644 index 0000000..35410ca --- /dev/null +++ b/homework7/.idea/.gitignore @@ -0,0 +1,8 @@ +# 默认忽略的文件 +/shelf/ +/workspace.xml +# 基于编辑器的 HTTP 客户端请求 +/httpRequests/ +# Datasource local storage ignored files +/dataSources/ +/dataSources.local.xml diff --git a/homework7/.idea/MarsCodeWorkspaceAppSettings.xml b/homework7/.idea/MarsCodeWorkspaceAppSettings.xml new file mode 100644 index 0000000..05ed8ba --- /dev/null +++ b/homework7/.idea/MarsCodeWorkspaceAppSettings.xml @@ -0,0 +1,6 @@ + + + + + \ No newline at end of file diff --git a/homework7/.idea/homework7.iml b/homework7/.idea/homework7.iml new file mode 100644 index 0000000..2051088 --- /dev/null +++ b/homework7/.idea/homework7.iml @@ -0,0 +1,8 @@ + + + + + + + + \ No newline at end of file diff --git a/homework7/.idea/inspectionProfiles/profiles_settings.xml b/homework7/.idea/inspectionProfiles/profiles_settings.xml new file mode 100644 index 0000000..105ce2d --- /dev/null +++ b/homework7/.idea/inspectionProfiles/profiles_settings.xml @@ -0,0 +1,6 @@ + + + + \ No newline at end of file diff --git a/homework7/.idea/misc.xml b/homework7/.idea/misc.xml new file mode 100644 index 0000000..6de7068 --- /dev/null +++ b/homework7/.idea/misc.xml @@ -0,0 +1,6 @@ + + + + + \ No newline at end of file diff --git a/homework7/.idea/modules.xml b/homework7/.idea/modules.xml new file mode 100644 index 0000000..f3eeb09 --- /dev/null +++ b/homework7/.idea/modules.xml @@ -0,0 +1,8 @@ + + + + + + + + \ No newline at end of file diff --git a/homework7/.idea/vcs.xml b/homework7/.idea/vcs.xml new file mode 100644 index 0000000..6c0b863 --- /dev/null +++ b/homework7/.idea/vcs.xml @@ -0,0 +1,6 @@ + + + + + + \ No newline at end of file diff --git a/homework7/main.py b/homework7/main.py new file mode 100644 index 0000000..68858d5 --- /dev/null +++ b/homework7/main.py @@ -0,0 +1,91 @@ +import numpy as np +# 在文件开头添加兼容性设置 +import tensorflow.compat.v1 as tf +tf.disable_v2_behavior() + +# 加载MNIST数据集，通过设置 one_hot=True 来使用独热编码标签 +# 独热编码：对于每个图片的标签 y，10 位中仅有一位的值为 1，其余的为 0。 +mnist = tf.keras.datasets.mnist +(x_train, y_train), (x_test, y_test) = mnist.load_data() + + +# 权重正态分布初始化函数 +def weight_variable(shape): + # 生成截断正态分布随机数,shape表示生成张量的维度，mean是均值(默认=0.0)，stddev是标准差。 + # 取值范围为 [ mean - 2 * stddev, mean + 2 * stddev ]，这里为[-0.2, 0.2] + initial = tf.truncated_normal(shape, stddev=0.1) + return tf.Variable(initial) + + +# 偏置量初始化函数 +def bias_variable(shape): + initial = tf.constant(0.1, shape=shape) # value=0.1, shape是张量的维度 + return tf.Variable(initial) + + +if __name__ == "__main__": + # 预处理数据 + x_train = x_train.reshape(-1, 784).astype('float32') / 255.0 + x_test = x_test.reshape(-1, 784).astype('float32') / 255.0 + y_train = tf.keras.utils.to_categorical(y_train, 10) + y_test = tf.keras.utils.to_categorical(y_test, 10) + + # 替换原来的mnist变量使用方式 + mnist_dataset = (x_train, y_train), (x_test, y_test) + + # 后续代码中所有mnist.train.xxx需要改为mnist_dataset[0][0]和mnist_dataset[0][1] + # 所有mnist.test.xxx需要改为mnist_dataset[1][0]和mnist_dataset[1][1] + print(x_train.shape[0]) # 输出训练集样本数 60000 + print(x_test.shape[0]) # 输出测试集样本数 10000 + + # 为训练数据集的输入 x 和标签 y 创建占位符 + x = tf.placeholder(tf.float32, [None, 784]) # 保持原样（已通过兼容性导入） + y = tf.placeholder(tf.float32, [None, 10]) + keep_prob = tf.placeholder(tf.float32) + + # 创建神经网络第1层，输入层，激活函数为relu + W_layer1 = weight_variable([784, 500]) + b_layer1 = bias_variable([500]) + h1 = tf.add(tf.matmul(x, W_layer1), b_layer1) # W * x + b + h1 = tf.nn.relu(h1) + # 创建神经网络第2层，隐藏层，激活函数为relu + W_layer2 = weight_variable([500, 1000]) + b_layer2 = bias_variable([1000]) + h2 = tf.add(tf.matmul(h1, W_layer2), b_layer2) # W * h1 + b，h1为第1层的输出 + h2 = tf.nn.relu(h2) + # 创建神经网络第3层，隐藏层，激活函数为relu + W_layer3 = weight_variable([1000, 300]) + b_layer3 = bias_variable([300]) + h3 = tf.add(tf.matmul(h2, W_layer3), b_layer3) # W * h2 + b，h2为第2层的输出 + h3 = tf.nn.relu(h3) + # 创建神经网络第4层，输出层，激活函数为softmax + W_layer4 = weight_variable([300, 10]) + b_layer4 = bias_variable([10]) + predict = tf.add(tf.matmul(h3, W_layer4), b_layer4) # W * h3 + b，h3为第3层的输出 + y_conv = tf.nn.softmax(tf.matmul(h3, W_layer4) + b_layer4) + # 计算交叉熵代价函数 + cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predict, labels=y)) + # 使用Adam下降算法优化交叉熵代价函数 + train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) + # 预测是否准确的结果存放在一个布尔型的列表中 + correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y, 1)) # argmax返回的矩阵行中的最大值的索引号 + # 求预测准确率 + accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float')) # cast将布尔型的数据转换成float型的数据；reduce_mean求平均值 + + # 初始化 + init_op = tf.global_variables_initializer() + + with tf.Session() as sess: + sess.run(init_op) + for i in range(550): # 训练样本为55000，分成550批，每批为100个样本 + # 手动实现next_batch + batch_index = np.random.randint(0, x_train.shape[0], 100) + batch = (x_train[batch_index], y_train[batch_index]) + train_accuracy = accuracy.eval(feed_dict={x: batch[0], y: batch[1], keep_prob: 1.0}) + test_accuracy = accuracy.eval(feed_dict={x: x_test, y: y_test}) + print('step %d, training accuracy %g, test accuracy %g' % (i, train_accuracy, test_accuracy)) + # 每一步迭代，都会加载100个训练样本，然后执行一次train_step，并通过feed_dict，用训练数据替代x和y张量占位符。 + sess.run(train_step, feed_dict={x: batch[0], y: batch[1], keep_prob: 0.5}) + # 显示最终在测试集上的准确率 + print( + 'test accuracy %g' % accuracy.eval(feed_dict={x: x_test, y: y_test, keep_prob: 1.0})) \ No newline at end of file