学卷积神经网络的理论的时候,我觉得自己看懂了,可是到了用代码来搭建一个卷积神经网络时,我发现自己有太多模糊的地方。这次还是基于MINIST数据集搭建一个卷积神经网络,首先给出一个基本的模型,然后再用Batch Norm、Dropout和早停对模型进行优化;在此过程中说明我在调试代码过程中遇到的一些问题和解决方法。
一、搭建基本的卷积神经网络
第一步:准备数据
在《Hands on Machine Learning with Scikit-Learn and TensorFlow》这本书上,用的是下面这一段代码来下载MINIST数据集。
from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("/tmp/data/")
用这种方式下载可能会报一个URLError的错误。大意是SSL证书验证失败,可以在前面加上下面那一段代码来取消SSL证书验证。
URLError:
import sslssl._create_default_https_context = ssl._create_unverified_context
然后运行后会出现一大堆的WARNING,但是不用担心,数据集还是能下载成功,而且还贴心地划分好了训练集、验证集和测试集,生成了batch,并reshape成了恰当的输入格式(比如训练集的维度已经是(55000, 784))。问题是下载太慢了,我失败了很多次,成功全靠运气。
我还是倾向于用tf.keras.datasets.mnist.load_data()来下载野生原始数据,然后自己动手划分数据、生成batch、整理成恰当的输入格式。
import tensorflow as tfimport numpy as npimport timefrom datetime import timedelta# 记录训练花费的时间def get_time_dif(start_time): end_time = time.time() time_dif = end_time - start_time #timedelta是用于对间隔进行规范化输出,间隔10秒的输出为:00:00:10 return timedelta(seconds=int(round(time_dif)))# 准备训练数据集、验证集和测试集,并生成小批量样本(X_train, y_train), (X_test, y_test) = tf.keras.datasets.mnist.load_data()# 对数据进行归一化,把训练集reshape成(60000,784)的维度X_train = X_train.astype(np.float32).reshape(-1, 28*28) / 255.0X_test = X_test.astype(np.float32).reshape(-1, 28*28) / 255.0y_train = y_train.astype(np.int32)y_test = y_test.astype(np.int32)# 划分训练集和验证集X_valid, X_train = X_train[:5000], X_train[5000:]y_valid, y_train = y_train[:5000], y_train[5000:]def shuffle_batch(X, y, batch_size): rnd_idx = np.random.permutation(len(X)) n_batches = len(X) // batch_size for batch_idx in np.array_split(rnd_idx, n_batches): X_batch, y_batch = X[batch_idx], y[batch_idx] yield X_batch, y_batch
第二步:配置参数
构建的网络有两个卷积层和一个全连接层,结构是:输入层—卷积层1—卷积层2—最大池化层—全连接层—输出层。卷积层又由卷积核与ReLU激活函数构成。
第一个卷积层有16个卷积核,尺寸为(3, 3),步幅为1,进行补零操作。第二个卷积层有32个卷积核,尺寸为(3,3),步幅为2,也进行补零。一般而言,越靠后的卷积层,输出的特征图要越多,而每个特征图的尺寸要越小,这就需要增加卷积核、增大卷积核尺寸和增大步幅。这样越往后就能提取到越高级的特征。
每个特征图上的神经元的参数(权重和偏置)是共享的,而不同特征图则有着不同的参数。每一个特征图都能提取出一个图片特征,这意味着特征图越多,提取到的图片特征也越多。
然后我们来看看相关的计算。假设卷积层的输入神经元个数为 n,卷积核大小为 m,步长为 s,输入神经元两端各填补p个零,那么该卷积层的输出神经元的个数为 (n-m+2p)/s + 1。由下面的参数可以知道,第1个卷积层输入神经元的数量为 n=28*28=784,m=3,s=1,由于padding=“SAME”,那么由 (784-3+2p)+1=784可知,p=1,也就是左右各补1个零。
可是在第2个卷积层,我却算出来补零的个数p不是整数,不知道是怎么进行后续操作的。
# 设定输入的高度、宽度、通道数height = 28width = 28channels = 1n_inputs = height * width# 设定卷积层特征图(过滤器)的个数,卷积核的尺寸、步幅conv1_fmaps = 16conv1_ksize = 3conv1_stride = 1conv1_pad = "SAME"conv2_fmaps = 32conv2_ksize = 3conv2_stride = 2conv2_pad = "SAME"# 最大池化层的特征图数量(通道数)pool3_fmaps = conv2_fmaps# 设定全连接层的神经元数量。n_fc1 = 32n_outputs = 10
第三步:构建卷积网络
下面的代码正是按照上面所说的网络结构去构建的,需要注意的地方有两点:一是最大池化时不要补零,因为池化的作用就是减少内存占用和参数数量;二是在输入到全连接层之前,要把所有特征图拉平成一个向量。
with tf.name_scope("inputs"): X = tf.placeholder(tf.float32, shape=[None, n_inputs], name="X") X_reshaped = tf.reshape(X, shape=[-1, height, width, channels]) y = tf.placeholder(tf.int32, shape=[None], name="y")conv1 = tf.layers.conv2d(X_reshaped, filters=conv1_fmaps, kernel_size=conv1_ksize, strides=conv1_stride, padding=conv1_pad, activation=tf.nn.relu, name="conv1")conv2 = tf.layers.conv2d(conv1, filters=conv2_fmaps, kernel_size=conv2_ksize, strides=conv2_stride, padding=conv2_pad, activation=tf.nn.relu, name="conv2")with tf.name_scope("pool3"): pool3 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID") # 把所有特征图拉平成一个向量,最大池化后特在图缩小为原来的1/16,所以由28*28变成了7*7 pool3_flat = tf.reshape(pool3, shape=[-1, pool3_fmaps * 7 * 7])with tf.name_scope("fc1"): fc1 = tf.layers.dense(pool3_flat, n_fc1, activation=tf.nn.relu, name="fc1")with tf.name_scope("output"): logits = tf.layers.dense(fc1, n_outputs, name="output") Y_proba = tf.nn.softmax(logits, name="Y_proba")with tf.name_scope("train"): xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y) loss = tf.reduce_mean(xentropy) optimizer = tf.train.AdamOptimizer() training_op = optimizer.minimize(loss)with tf.name_scope("eval"): correct = tf.nn.in_top_k(logits, y, 1) accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
第四步:训练和评估模型
训练和评估阶段最大的问题就是在卷积层可能存在内存溢出,尤其是评估和测试时。训练时batch-size=100,问题不大,而验证集的样本数为5000,测试集的样本数为10000,在计算时是非常消耗内存的。我在测试时,就出现了如下的错误:
ResourceExhaustedError: OOM when allocating tensor with shape[10000,16,29,29]...
OOM意思就是“ Out of Memorry”,这段错误是指在测试阶段内存溢出了。我的GPU是GTX960M,显卡内存是2G,实际训练模型时可用的大概是1.65G,还是比较小。
遇到这种问题,有几种解决办法:一种是让模型简单点,比如减少卷积层的特征图数量,增加步幅,减少卷积层的数量,但是这一般会让模型的性能下降;第二种方法是把32位的浮点数改为16位的;第三种方法是在评估和测试时也进行小批量操作。
让模型变得简单会减低模型的性能,我试了,的确如此,因此我选择了第三种方法,在评估和测试时,把数据按每批次1000个样本输入,然后求平均值。最终的验证精度为98.74%。
with tf.name_scope("init_and_save"): init = tf.global_variables_initializer() saver = tf.train.Saver() n_epochs = 10 batch_size = 100 with tf.Session() as sess: init.run() start_time = time.time() for epoch in range(n_epochs): for X_batch, y_batch in shuffle_batch(X_train,y_train,batch_size): sess.run(training_op, feed_dict={X: X_batch, y: y_batch}) acc_train = accuracy.eval(feed_dict={X: X_batch, y: y_batch}) if epoch % 2 == 0 or epoch == 9: # 每次输入1000个样本进行评估,然后求平均值 acc_val = [] for i in range(len(X_valid)//1000): acc_val.append(accuracy.eval(feed_dict={X: X_valid[i*1000:(i+1)*1000], y: y_valid[i*1000:(i+1)*1000]})) acc_val = np.mean(acc_val) print('Epoch:{0:>4}, Train accuracy:{1:>7.2%},Validate accuracy:{2:7.2%}'.format(epoch,acc_train, acc_val)) time_dif = get_time_dif(start_time) print("\nTime usage:", time_dif) acc_test = [] # 每次输入1000个样本进行测试,再求平均值 for i in range(len(X_test)//1000): acc_test.append(accuracy.eval(feed_dict={X: X_test[i*1000:(i+1)*1000], y: y_test[i*1000:(i+1)*1000]})) acc_test = np.mean(acc_test) print("\nTest_accuracy:{0:>7.2%}".format(acc_test))
Epoch: 0, Train accuracy: 98.00%,Validate accuracy: 97.12%Epoch: 2, Train accuracy: 97.00%,Validate accuracy: 98.34%Epoch: 4, Train accuracy:100.00%,Validate accuracy: 98.62%Epoch: 6, Train accuracy:100.00%,Validate accuracy: 98.84%Epoch: 8, Train accuracy: 99.00%,Validate accuracy: 98.68%Epoch: 9, Train accuracy:100.00%,Validate accuracy: 98.86%Time usage: 0:01:02Test_accuracy: 98.68%
二、用Batch Norm、Dropout和早停优化卷积神经网络
参考的这本书里用Dropout和早停来优化卷积神经网络的基本模型,没有用Batch Norm来优化。我觉得作者实现早停的代码太复杂了,推荐用我的这个代码来实现,清晰明了。
关于在卷积神经网络中运用Batch Norm的代码我暂时没找到,只能凭自己的理解来实现。Batch Norm在哪些层用呢?我觉得在卷积层和全连接层(包括输出层)用,在池化层就不用了,因为内部协变量偏移问题应该主要源自于层与层之间的非线性变换,而池化层的输出值并没有做非线性激活,因此在之后的全连接层做Batch Norm就行。
Dropout运用在池化层和全连接层,丢弃率分别为0.25和0.5,注意是按照Batch Norm—SELU函数激活—Dropout的顺序来进行。
同时将第2个卷积层的卷积步幅设置为1,以获得尺寸更大的特征图和更多参数。
设置迭代轮次为20,batch size = 100,做Batch Norm 时因为要求每个小批量的均值和方差,因此batch size 可以稍微设置得大一些。如果2000步以后验证精度仍然没有提升,那就中止训练。
结果,模型在第18轮、第9921步中止了训练,最好的验证精度为99.22%,测试精度为98.94%。
import tensorflow as tfimport numpy as npimport timefrom datetime import timedeltafrom functools import partial# 记录训练花费的时间def get_time_dif(start_time): end_time = time.time() time_dif = end_time - start_time #timedelta是用于对间隔进行规范化输出,间隔10秒的输出为:00:00:10 return timedelta(seconds=int(round(time_dif)))# 准备训练数据集、验证集和测试集,并生成小批量样本(X_train, y_train), (X_test, y_test) = tf.keras.datasets.mnist.load_data()X_train = X_train.astype(np.float32).reshape(-1, 28*28) / 255.0X_test = X_test.astype(np.float32).reshape(-1, 28*28) / 255.0y_train = y_train.astype(np.int32)y_test = y_test.astype(np.int32)X_valid, X_train = X_train[:5000], X_train[5000:]y_valid, y_train = y_train[:5000], y_train[5000:]def shuffle_batch(X, y, batch_size): rnd_idx = np.random.permutation(len(X)) n_batches = len(X) // batch_size for batch_idx in np.array_split(rnd_idx, n_batches): X_batch, y_batch = X[batch_idx], y[batch_idx] yield X_batch, y_batchheight = 28width = 28channels = 1n_inputs = height * width# 第一个卷积层有16个卷积核# 卷积核的大小为(3,3)# 步幅为1# 通过补零让输入与输出的维度相同conv1_fmaps = 16conv1_ksize = 3conv1_stride = 1conv1_pad = "SAME"conv2_fmaps = 32conv2_ksize = 3conv2_stride = 1conv2_pad = "SAME"# 在池化层丢弃25%的神经元conv2_dropout_rate = 0.25pool3_fmaps = conv2_fmapsn_fc1 = 32# 在全连接层丢弃50%的神经元fc1_dropout_rate = 0.5n_outputs = 10with tf.name_scope("inputs"): X = tf.placeholder(tf.float32, shape=[None, n_inputs], name="X") X_reshaped = tf.reshape(X, shape=[-1, height, width, channels]) y = tf.placeholder(tf.int32, shape=[None], name="y") training = tf.placeholder_with_default(False, shape=[], name='training')# 构建一个batch norm层,便于复用。用移动平均求全局的样本均值和方差,动量参数取0.9my_batch_norm_layer = partial(tf.layers.batch_normalization, training=training, momentum=0.9)with tf.name_scope("conv"): # batch norm之后在激活,所以这里不设定激活函数 conv1 = tf.layers.conv2d(X_reshaped, filters=conv1_fmaps, kernel_size=conv1_ksize, strides=conv1_stride, padding=conv1_pad, activation=None, name="conv1") # 进行batch norm之后,再激活 batch_norm1 = tf.nn.selu(my_batch_norm_layer(conv1)) conv2 = tf.layers.conv2d(batch_norm1, filters=conv2_fmaps, kernel_size=conv2_ksize, strides=conv2_stride, padding=conv2_pad, activation=None, name="conv2") batch_norm2 = tf.nn.selu(my_batch_norm_layer(conv2)) with tf.name_scope("pool3"): pool3 = tf.nn.max_pool(batch_norm2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID") # 把特征图拉平成一个向量 pool3_flat = tf.reshape(pool3, shape=[-1, pool3_fmaps * 14 * 14]) # 丢弃25%的神经元 pool3_flat_drop = tf.layers.dropout(pool3_flat, conv2_dropout_rate, training=training) with tf.name_scope("fc1"): fc1 = tf.layers.dense(pool3_flat_drop, n_fc1, activation=None, name="fc1") # 在全连接层进行batch norm,然后激活 batch_norm4 = tf.nn.selu(my_batch_norm_layer(fc1)) # 丢弃50%的神经元 fc1_drop = tf.layers.dropout(batch_norm4, fc1_dropout_rate, training=training) with tf.name_scope("output"): logits = tf.layers.dense(fc1_drop, n_outputs, name="output") logits_batch_norm = my_batch_norm_layer(logits) Y_proba = tf.nn.softmax(logits_batch_norm, name="Y_proba")with tf.name_scope("loss_and_train"): learning_rate = 0.01 xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits_batch_norm, labels=y) loss = tf.reduce_mean(xentropy) optimizer = tf.train.AdamOptimizer(learning_rate) # 这是需要额外更新batch norm的参数 extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) # 模型参数的优化依赖与batch norm参数的更新 with tf.control_dependencies(extra_update_ops): training_op = optimizer.minimize(loss)with tf.name_scope("eval"): correct = tf.nn.in_top_k(logits, y, 1) accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))with tf.name_scope("init_and_save"): init = tf.global_variables_initializer() saver = tf.train.Saver() n_epochs = 20batch_size = 100with tf.Session() as sess: init.run() start_time = time.time() # 记录总迭代步数,一个batch算一步 # 记录最好的验证精度 # 记录上一次验证结果提升时是第几步。 # 如果迭代2000步后结果还没有提升就中止训练。 total_batch = 0 best_acc_val = 0.0 last_improved = 0 require_improvement = 2000 flag = False for epoch in range(n_epochs): for X_batch, y_batch in shuffle_batch(X_train, y_train, batch_size): sess.run(training_op, feed_dict={training:True, X: X_batch, y: y_batch}) # 每次迭代10步就验证一次 if total_batch % 10 == 0: acc_batch = accuracy.eval(feed_dict={X: X_batch, y: y_batch}) # 每次输入1000个样本进行评估,然后求平均值 acc_val = [] for i in range(len(X_valid)//1000): acc_val.append(accuracy.eval(feed_dict={X: X_valid[i*1000:(i+1)*1000], y: y_valid[i*1000:(i+1)*1000]})) acc_val = np.mean(acc_val) # 如果验证精度提升了,就替换为最好的结果,并保存模型 if acc_val > best_acc_val: best_acc_val = acc_val last_improved = total_batch save_path = saver.save(sess, "./my_model_CNN_stop.ckpt") improved_str = 'improved!' else: improved_str = '' # 记录训练时间,并格式化输出验证结果,如果提升了,会在后面提示:improved! time_dif = get_time_dif(start_time) msg = 'Epoch:{0:>4}, Iter: {1:>6}, Acc_Train: {2:>7.2%}, Acc_Val: {3:>7.2%}, Time: {4} {5}' print(msg.format(epoch, total_batch, acc_batch, acc_val, time_dif, improved_str)) # 记录总迭代步数 total_batch += 1 # 如果2000步以后还没提升,就中止训练。 if total_batch - last_improved > require_improvement: print("Early stopping in ",total_batch," step! And the best validation accuracy is ",best_acc_val, '.') # 跳出这个轮次的循环 flag = True break # 跳出所有训练轮次的循环 if flag: break with tf.Session() as sess: saver.restore(sess, "./my_model_CNN_stop.ckpt") # 每次输入1000个样本进行测试,再求平均值 acc_test = [] for i in range(len(X_test)//1000): acc_test.append(accuracy.eval(feed_dict={X: X_test[i*1000:(i+1)*1000], y: y_test[i*1000:(i+1)*1000]})) acc_test = np.mean(acc_test) print("\nTest_accuracy:{0:>7.2%}".format(acc_test))
Early stopping in 9921 step! And the best validation accuracy is 0.9922 .INFO:tensorflow:Restoring parameters from ./my_model_CNN_stop.ckptTest_accuracy: 98.94%
参考资料:
《Hands on Machine Learning with Scikit-Learn and TensorFlow》