『TensorFlow』单&双隐藏层自编码器设计
计算图设计
很简单的实践,
- 多了个隐藏层
- 没有上节的高斯噪声
- 网络写法由上节的面向对象改为了函数式编程,
其他没有特别需要注意的,实现如下:
import numpy as np import matplotlib.pyplot as plt import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' learning_rate = 0.01 # 学习率 training_epochs = 20 # 训练轮数,1轮等于n_samples/batch_size batch_size = 128 # batch容量 display_step = 1 # 展示间隔 example_to_show = 10 # 展示图像数目 n_hidden_units = 256 n_input_units = 784 n_output_units = n_input_units def WeightsVariable(n_in, n_out, name_str): return tf.Variable(tf.random_normal([n_in, n_out]), dtype=tf.float32, name=name_str) def biasesVariable(n_out, name_str): return tf.Variable(tf.random_normal([n_out]), dtype=tf.float32, name=name_str) def encoder(x_origin, activate_func=tf.nn.sigmoid): with tf.name_scope('Layer'): Weights = WeightsVariable(n_input_units, n_hidden_units, 'Weights') biases = biasesVariable(n_hidden_units, 'biases') x_code = activate_func(tf.add(tf.matmul(x_origin, Weights), biases)) return x_code def decode(x_code, activate_func=tf.nn.sigmoid): with tf.name_scope('Layer'): Weights = WeightsVariable(n_hidden_units, n_output_units, 'Weights') biases = biasesVariable(n_output_units, 'biases') x_decode = activate_func(tf.add(tf.matmul(x_code, Weights), biases)) return x_decode with tf.Graph().as_default(): with tf.name_scope('Input'): X_input = tf.placeholder(tf.float32, [None, n_input_units]) with tf.name_scope('Encode'): X_code = encoder(X_input) with tf.name_scope('decode'): X_decode = decode(X_code) with tf.name_scope('loss'): loss = tf.reduce_mean(tf.pow(X_input - X_decode, 2)) with tf.name_scope('train'): Optimizer = tf.train.RMSPropOptimizer(learning_rate) train = Optimizer.minimize(loss) init = tf.global_variables_initializer() # 因为使用了tf.Graph.as_default()上下文环境 # 所以下面的记录必须放在上下文里面,否则记录下来的图是空的(get不到上面的default) writer = tf.summary.FileWriter(logdir='logs', graph=tf.get_default_graph()) writer.flush()
计算图:
训练程序
import numpy as np import matplotlib.pyplot as plt import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' learning_rate = 0.01 # 学习率 training_epochs = 20 # 训练轮数,1轮等于n_samples/batch_size batch_size = 128 # batch容量 display_step = 1 # 展示间隔 example_to_show = 10 # 展示图像数目 n_hidden_units = 256 n_input_units = 784 n_output_units = n_input_units def WeightsVariable(n_in, n_out, name_str): return tf.Variable(tf.random_normal([n_in, n_out]), dtype=tf.float32, name=name_str) def biasesVariable(n_out, name_str): return tf.Variable(tf.random_normal([n_out]), dtype=tf.float32, name=name_str) def encoder(x_origin, activate_func=tf.nn.sigmoid): with tf.name_scope('Layer'): Weights = WeightsVariable(n_input_units, n_hidden_units, 'Weights') biases = biasesVariable(n_hidden_units, 'biases') x_code = activate_func(tf.add(tf.matmul(x_origin, Weights), biases)) return x_code def decode(x_code, activate_func=tf.nn.sigmoid): with tf.name_scope('Layer'): Weights = WeightsVariable(n_hidden_units, n_output_units, 'Weights') biases = biasesVariable(n_output_units, 'biases') x_decode = activate_func(tf.add(tf.matmul(x_code, Weights), biases)) return x_decode with tf.Graph().as_default(): with tf.name_scope('Input'): X_input = tf.placeholder(tf.float32, [None, n_input_units]) with tf.name_scope('Encode'): X_code = encoder(X_input) with tf.name_scope('decode'): X_decode = decode(X_code) with tf.name_scope('loss'): loss = tf.reduce_mean(tf.pow(X_input - X_decode, 2)) with tf.name_scope('train'): Optimizer = tf.train.RMSPropOptimizer(learning_rate) train = Optimizer.minimize(loss) init = tf.global_variables_initializer() # 因为使用了tf.Graph.as_default()上下文环境 # 所以下面的记录必须放在上下文里面,否则记录下来的图是空的(get不到上面的default) writer = tf.summary.FileWriter(logdir='logs', graph=tf.get_default_graph()) writer.flush() mnist = input_data.read_data_sets('../Mnist_data/', one_hot=True) with tf.Session() as sess: sess.run(init) total_batch = int(mnist.train.num_examples / batch_size) for epoch in range(training_epochs): for i in range(total_batch): batch_xs, batch_ys = mnist.train.next_batch(batch_size) _, Loss = sess.run([train, loss], feed_dict={X_input: batch_xs}) Loss = sess.run(loss, feed_dict={X_input: batch_xs}) if epoch % display_step == 0: print('Epoch: %04d' % (epoch + 1), 'loss= ', '{:.9f}'.format(Loss)) writer.close() print('训练完毕!') '''比较输入和输出的图像''' # 输出图像获取 reconstructions = sess.run(X_decode, feed_dict={X_input: mnist.test.images[:example_to_show]}) # 画布建立 f, a = plt.subplots(2, 10, figsize=(10, 2)) for i in range(example_to_show): a[0][i].imshow(np.reshape(mnist.test.images[i], (28, 28))) a[1][i].imshow(np.reshape(reconstructions[i], (28, 28))) f.show() # 渲染图像 plt.draw() # 刷新图像 # plt.waitforbuttonpress()
debug一上午的收获:接受sess.run输出的变量名不要和tensor节点的变量名重复,会出错的... ...好低级的错误。mmdz
比较图像一部分之前没做过,介绍了matplotlib.pyplot的花式用法,
原来plt.subplots()是会返回 画布句柄 & 子图集合 句柄的,子图集合句柄可以像数组一样调用子图
pyplot是有show()和draw()两个方法的,show是展示出画布,draw会刷新原图,可以交互的修改画布
waitforbuttonpress()监听键盘按键如果用户按的是键盘,返回True,如果是其他(如鼠标单击),则返回False
另,发现用surface写程序其实还挺带感... ...
输出图像如下:
双隐藏层版本
import numpy as np import matplotlib.pyplot as plt import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' batch_size = 128 # batch容量 display_step = 1 # 展示间隔 learning_rate = 0.01 # 学习率 training_epochs = 20 # 训练轮数,1轮等于n_samples/batch_size example_to_show = 10 # 展示图像数目 n_hidden1_units = 256 # 第一隐藏层 n_hidden2_units = 128 # 第二隐藏层 n_input_units = 784 n_output_units = n_input_units def WeightsVariable(n_in, n_out, name_str): return tf.Variable(tf.random_normal([n_in, n_out]), dtype=tf.float32, name=name_str) def biasesVariable(n_out, name_str): return tf.Variable(tf.random_normal([n_out]), dtype=tf.float32, name=name_str) def encoder(x_origin, activate_func=tf.nn.sigmoid): with tf.name_scope('Layer1'): Weights = WeightsVariable(n_input_units, n_hidden1_units, 'Weights') biases = biasesVariable(n_hidden1_units, 'biases') x_code1 = activate_func(tf.add(tf.matmul(x_origin, Weights), biases)) with tf.name_scope('Layer2'): Weights = WeightsVariable(n_hidden1_units, n_hidden2_units, 'Weights') biases = biasesVariable(n_hidden2_units, 'biases') x_code2 = activate_func(tf.add(tf.matmul(x_code1, Weights), biases)) return x_code2 def decode(x_code, activate_func=tf.nn.sigmoid): with tf.name_scope('Layer1'): Weights = WeightsVariable(n_hidden2_units, n_hidden1_units, 'Weights') biases = biasesVariable(n_hidden1_units, 'biases') x_decode1 = activate_func(tf.add(tf.matmul(x_code, Weights), biases)) with tf.name_scope('Layer2'): Weights = WeightsVariable(n_hidden1_units, n_output_units, 'Weights') biases = biasesVariable(n_output_units, 'biases') x_decode2 = activate_func(tf.add(tf.matmul(x_decode1, Weights), biases)) return x_decode2 with tf.Graph().as_default(): with tf.name_scope('Input'): X_input = tf.placeholder(tf.float32, [None, n_input_units]) with tf.name_scope('Encode'): X_code = encoder(X_input) with tf.name_scope('decode'): X_decode = decode(X_code) with tf.name_scope('loss'): loss = tf.reduce_mean(tf.pow(X_input - X_decode, 2)) with tf.name_scope('train'): Optimizer = tf.train.RMSPropOptimizer(learning_rate) train = Optimizer.minimize(loss) init = tf.global_variables_initializer() # 因为使用了tf.Graph.as_default()上下文环境 # 所以下面的记录必须放在上下文里面,否则记录下来的图是空的(get不到上面的default) writer = tf.summary.FileWriter(logdir='logs', graph=tf.get_default_graph()) writer.flush() mnist = input_data.read_data_sets('../Mnist_data/', one_hot=True) with tf.Session() as sess: sess.run(init) total_batch = int(mnist.train.num_examples / batch_size) for epoch in range(training_epochs): for i in range(total_batch): batch_xs, batch_ys = mnist.train.next_batch(batch_size) _, Loss = sess.run([train, loss], feed_dict={X_input: batch_xs}) Loss = sess.run(loss, feed_dict={X_input: batch_xs}) if epoch % display_step == 0: print('Epoch: %04d' % (epoch + 1), 'loss= ', '{:.9f}'.format(Loss)) writer.close() print('训练完毕!') '''比较输入和输出的图像''' # 输出图像获取 reconstructions = sess.run(X_decode, feed_dict={X_input: mnist.test.images[:example_to_show]}) # 画布建立 f, a = plt.subplots(2, 10, figsize=(10, 2)) for i in range(example_to_show): a[0][i].imshow(np.reshape(mnist.test.images[i], (28, 28))) a[1][i].imshow(np.reshape(reconstructions[i], (28, 28))) f.show() # 渲染图像 plt.draw() # 刷新图像 # plt.waitforbuttonpress()
输出图像如下:
由于压缩到128个节点损失信息过多,所以结果不如之前单层的好。
有意思的是我们把256的那层改成128(也就是双128)后,结果反而比上面的要好:
但是仍然比不上单隐藏层,数据比较简单时候复杂网络效果可能不那么好(loss值我没有截取,但实际上是这样,虽然不同网络loss直接比较没什么意义),当然,也有可能是复杂网络没收敛的结果。
可视化双隐藏层自编码器
import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' batch_size = 128 # batch容量 display_step = 1 # 展示间隔 learning_rate = 0.01 # 学习率 training_epochs = 20 # 训练轮数,1轮等于n_samples/batch_size example_to_show = 10 # 展示图像数目 n_hidden1_units = 256 # 第一隐藏层 n_hidden2_units = 128 # 第二隐藏层 n_input_units = 784 n_output_units = n_input_units def variable_summaries(var): #<--- """ 可视化变量全部相关参数 :param var: :return: """ with tf.name_scope('summaries'): mean = tf.reduce_mean(var) tf.summary.histogram('mean', mean) with tf.name_scope('stddev'): stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean))) tf.summary.scalar('stddev', stddev) # 注意,这是标量 tf.summary.scalar('max', tf.reduce_max(var)) tf.summary.scalar('min', tf.reduce_min(var)) tf.summary.histogram('histogram', var) def WeightsVariable(n_in,n_out,name_str): return tf.Variable(tf.random_normal([n_in,n_out]),dtype=tf.float32,name=name_str) def biasesVariable(n_out,name_str): return tf.Variable(tf.random_normal([n_out]),dtype=tf.float32,name=name_str) def encoder(x_origin,activate_func=tf.nn.sigmoid): with tf.name_scope('Layer1'): Weights = WeightsVariable(n_input_units,n_hidden1_units,'Weights') biases = biasesVariable(n_hidden1_units,'biases') x_code1 = activate_func(tf.add(tf.matmul(x_origin,Weights),biases)) variable_summaries(Weights) #<--- variable_summaries(biases) #<--- with tf.name_scope('Layer2'): Weights = WeightsVariable(n_hidden1_units,n_hidden2_units,'Weights') biases = biasesVariable(n_hidden2_units,'biases') x_code2 = activate_func(tf.add(tf.matmul(x_code1,Weights),biases)) variable_summaries(Weights) #<--- variable_summaries(biases) #<--- return x_code2 def decode(x_code,activate_func=tf.nn.sigmoid): with tf.name_scope('Layer1'): Weights = WeightsVariable(n_hidden2_units,n_hidden1_units,'Weights') biases = biasesVariable(n_hidden1_units,'biases') x_decode1 = activate_func(tf.add(tf.matmul(x_code,Weights),biases)) variable_summaries(Weights) #<--- variable_summaries(biases) #<--- with tf.name_scope('Layer2'): Weights = WeightsVariable(n_hidden1_units,n_output_units,'Weights') biases = biasesVariable(n_output_units,'biases') x_decode2 = activate_func(tf.add(tf.matmul(x_decode1,Weights),biases)) variable_summaries(Weights) #<--- variable_summaries(biases) #<--- return x_decode2 with tf.Graph().as_default(): with tf.name_scope('Input'): X_input = tf.placeholder(tf.float32,[None,n_input_units]) with tf.name_scope('Encode'): X_code = encoder(X_input) with tf.name_scope('decode'): X_decode = decode(X_code) with tf.name_scope('loss'): loss = tf.reduce_mean(tf.pow(X_input - X_decode,2)) with tf.name_scope('train'): Optimizer = tf.train.RMSPropOptimizer(learning_rate) train = Optimizer.minimize(loss) # 标量汇总 with tf.name_scope('LossSummary'): tf.summary.scalar('loss',loss) tf.summary.scalar('learning_rate',learning_rate) # 图像展示 with tf.name_scope('ImageSummary'): image_original = tf.reshape(X_input,[-1, 28, 28, 1]) image_reconstruction = tf.reshape(X_decode, [-1, 28, 28, 1]) tf.summary.image('image_original', image_original, 9) tf.summary.image('image_recinstruction', image_reconstruction, 9) # 汇总 merged_summary = tf.summary.merge_all() init = tf.global_variables_initializer() writer = tf.summary.FileWriter(logdir='logs', graph=tf.get_default_graph()) writer.flush() mnist = input_data.read_data_sets('../Mnist_data/', one_hot=True) with tf.Session() as sess: sess.run(init) total_batch = int(mnist.train.num_examples / batch_size) for epoch in range(training_epochs): for i in range(total_batch): batch_xs,batch_ys = mnist.train.next_batch(batch_size) _,Loss = sess.run([train,loss],feed_dict={X_input: batch_xs}) Loss = sess.run(loss,feed_dict={X_input: batch_xs}) if epoch % display_step == 0: print('Epoch: %04d' % (epoch + 1),'loss= ','{:.9f}'.format(Loss)) summary_str = sess.run(merged_summary,feed_dict={X_input: batch_xs}) #<--- writer.add_summary(summary_str,epoch) #<--- writer.flush() #<--- writer.close() print('训练完毕!')
几个有意思的发现,
使用之前的图像输出方式时,win下matplotlib.pyplot的绘画框会立即退出,所以要使用 plt.waitforbuttonpress() 命令。
win下使用plt绘画色彩和linux不一样,效果如下:
输出图如下:
对比图像如下(截自tensorboard):