4.2.4 VGG-ResNet实战
VGGNET实战
VGGNET的思想就是加深神经网络层次,多使用3*3的卷积核替换5*5的
这里我们就不使用1*1的卷积核了
我们可以在之前的卷积神经网络基础上复用数据处理和测试的代码
只修改卷积层部分
# conv1:神经元图,feature map,输出图像conv1_1 = tf.layers.conv2d(x_image, 32, # output channel number (3,3), # kernal size padding = 'same', # same 代表输出图像的大小没有变化,valid 代表不做padding activation = tf.nn.relu, name = 'conv1_1' ) conv1_2 = tf.layers.conv2d(conv1_1, 32, # output channel number (3,3), # kernal size padding = 'same', # same 代表输出图像的大小没有变化,valid 代表不做padding activation = tf.nn.relu, name = 'conv1_2' )# 16*16pooling1 = tf.layers.max_pooling2d(conv1_2, (2, 2), # kernal size (2, 2), # stride name = 'pool1' # name为了给这一层做一个命名,这样会让图打印出来的时候会是一个有意义的图 ) conv2_1 = tf.layers.conv2d(pooling1, 32, # output channel number (3,3), # kernal size padding = 'same', # same 代表输出图像的大小没有变化,valid 代表不做padding activation = tf.nn.relu, name = 'conv2_1' ) conv2_2 = tf.layers.conv2d(conv2_1, 32, # output channel number (3,3), # kernal size padding = 'same', # same 代表输出图像的大小没有变化,valid 代表不做padding activation = tf.nn.relu, name = 'conv2_2' )# 8*8pooling2 = tf.layers.max_pooling2d(conv2_2, (2, 2), # kernal size (2, 2), # stride name = 'pool2' # name为了给这一层做一个命名,这样会让图打印出来的时候会是一个有意义的图 ) conv3_1 = tf.layers.conv2d(pooling2, 32, # output channel number (3,3), # kernal size padding = 'same', # same 代表输出图像的大小没有变化,valid 代表不做padding activation = tf.nn.relu, name = 'conv3_1' ) conv3_2 = tf.layers.conv2d(conv3_1, 32, # output channel number (3,3), # kernal size padding = 'same', # same 代表输出图像的大小没有变化,valid 代表不做padding activation = tf.nn.relu, name = 'conv3_2' )# 4*4*32pooling3 = tf.layers.max_pooling2d(conv3_2, (2, 2), # kernal size (2, 2), # stride name = 'pool3' # name为了给这一层做一个命名,这样会让图打印出来的时候会是一个有意义的图 )
训练10000次 可以达到百分之70的准确率
[Train] Step: 500, loss: 1.92473, acc: 0.45000[Train] Step: 1000, loss: 1.49288, acc: 0.35000[Train] Step: 1500, loss: 1.30839, acc: 0.55000[Train] Step: 2000, loss: 1.41633, acc: 0.40000[Train] Step: 2500, loss: 1.10951, acc: 0.60000[Train] Step: 3000, loss: 1.15743, acc: 0.65000[Train] Step: 3500, loss: 0.93834, acc: 0.70000[Train] Step: 4000, loss: 0.76699, acc: 0.80000[Train] Step: 4500, loss: 0.71109, acc: 0.70000[Train] Step: 5000, loss: 0.75763, acc: 0.75000(10000, 3072) (10000,)[Test ] Step: 5000, acc: 0.67500[Train] Step: 5500, loss: 0.98661, acc: 0.65000[Train] Step: 6000, loss: 1.43098, acc: 0.50000[Train] Step: 6500, loss: 0.86575, acc: 0.70000[Train] Step: 7000, loss: 0.80474, acc: 0.65000[Train] Step: 7500, loss: 0.60132, acc: 0.85000[Train] Step: 8000, loss: 0.66683, acc: 0.80000[Train] Step: 8500, loss: 0.56874, acc: 0.85000[Train] Step: 9000, loss: 0.68185, acc: 0.70000[Train] Step: 9500, loss: 0.83302, acc: 0.70000[Train] Step: 10000, loss: 0.87228, acc: 0.70000(10000, 3072) (10000,)[Test ] Step: 10000, acc: 0.72700
RESNET实战
先来回顾一下RESNET的网络结构
image.png
RESNET是先经过了一个卷积层,又经过了一个池化层,然后再经过若干个残差连接块
这里每经过一个残差连接块以后,可能会经过一个降采样的过程
所谓降采样就是之前的maxpooling或者卷积层的步长等于2
在上面的ResNet中,经过了四次降采样的过程,但是由于我们的实战使用的图片是32*32的本身就比较小,所以不会经过太多的降采样,也不会首先经过maxpooling层
在降采样的过程中可能会出现的一个问题是:残差有两部分组成,一部分是卷积操作,一部分是恒等变换,如果卷及操作降采样了,那么会导致两部分的维度不一样,这时候的矩阵加法会出问题。所以这个时候需要额外进行一个操作,就是如果卷积做了降采样,那么恒等变化也要做一次降采样,这个操作使用maxpooling来做。
image.png
先定义残差块的实现方法
""" x是输入数据,output_channel 是输出通道数 为了避免降采样带来的数据损失,我们会在降采样的时候讲output_channel翻倍 所以这里如果output_channel是input_channel的二倍,则说明需要降采样 """def residual_block(x, output_channel): """residual connection implementation""" input_channel = x.get_shape().as_list()[-1] if input_channel * 2 == output_channel: increase_dim = True strides = (2, 2) elif input_channel == output_channel: increase_dim = False strides = (1, 1) else: raise Exception("input channel can't match output channel") conv1 = tf.layers.conv2d(x, output_channel, (3,3), strides = strides, padding = 'same', activation = tf.nn.relu, name = 'conv1') conv2 = tf.layers.conv2d(conv1, output_channel, (3,3), strides = (1,1), padding = 'same', activation = tf.nn.relu, name = 'conv2') # 处理另一个分支(恒等变换) if increase_dim: # 需要降采样 # [None,image_width,image_height,channel] -> [,,,channel*2] pooled_x = tf.layers.average_pooling2d(x, (2,2), # pooling 核 (2,2), # strides strides = pooling 不重叠 padding = 'valid' # 这里图像大小是32*32,都能除尽,padding是什么没有关系 ) # average_pooling2d使得图的大小变化了,但是output_channel还是不匹配,下面修改output_channel padded_x = tf.pad(pooled_x, [[0,0], [0,0], [0,0], [input_channel // 2,input_channel //2]]) else: padded_x = x output_x = conv2 + padded_x return output_x
然后定义残差网络
先使用一个卷积层,然后循环创建残差块,最后跟一个全局的池化,然后是全连接到输出
全局的池化和普通的池化一样,只不过他的size和图像的width,height一样大,这样一个图像的输出就是一个数
def res_net(x, num_residual_blocks, num_filter_base, class_num): """residual network implementation""" """ Args: - x: 输入数据 - num_residual_blocks: 残差链接块数 eg: [3,4,6,3] - num_filter_base: 最初的通道数目 - class_num: 类别数目 """ # 需要做多少次降采样 num_subsampling = len(num_residual_blocks) layers = [] # [None,image_width,image_height,channel] -> [image_width,image_height,channel] # kernal size:image_width,image_height input_size = x.get_shape().as_list()[1:] with tf.variable_scope('conv0'): conv0 = tf.layers.conv2d(x, num_filter_base, (3,3), strides = (1,1), activation = tf.nn.relu, padding = 'same', name = 'conv0') layers.append(conv0) # eg: num_subsampling = 4 ,sample_id = [1,2,3,4] for sample_id in range(num_subsampling): for i in range(num_residual_blocks[sample_id]): with tf.variable_scope("conv%d_%d" % (sample_id, i)): conv = residual_block( layers[-1], num_filter_base * (2 ** sample_id)) # 每次翻倍 layers.append(conv) multiplier = 2 ** (num_subsampling - 1) assert layers[-1].get_shape().as_list()[1:] \ == [input_size[0] / multiplier, input_size[1] / multiplier, num_filter_base * multiplier] with tf.variable_scope('fc'): # layers[-1].shape : [None, width, height, channel] global_pool = tf.reduce_mean(layers[-1], [1, 2]) # pooling logits = tf.layers.dense(global_pool, class_num) # 全连接 layers.append(logits) return layers[-1]
然后使用残差网络
x = tf.placeholder(tf.float32, [None, 3072]) y = tf.placeholder(tf.int64, [None])# 将向量变成具有三通道的图片的格式x_image = tf.reshape(x, [-1,3,32,32])# 32*32x_image = tf.transpose(x_image, perm = [0, 2, 3, 1]) y_ = res_net(x_image, [2,3,2], 32, 10)# 交叉熵loss = tf.losses.sparse_softmax_cross_entropy(labels=y, logits=y_)# y_-> softmax# y -> one_hot# loss = ylogy_# boolpredict = tf.argmax(y_, 1)# [1,0,1,1,1,0,0,0]correct_prediction = tf.equal(predict, y) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float64))with tf.name_scope('train_op'): train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)
这里训练的结构过7000次百分之67.之所以比VGG低,是因为很多优化没有用。优化后的残差网络在cifar10上可以达到94%的准确率
作者:Meet相识_bfa5
链接:https://www.jianshu.com/p/14e95c300562