本文介绍了一种使用TensorFlow框架在CIFAR-10数据集上训练深度学习模型的方法，并详细展示了如何通过TFRecords读取数据、实现BatchNormalization层、构建卷积神经网络以及进行模型保存和测试的过程。
模型保存方法：
saver = tf.train.Saver(tf.global_variables())
下文中在cifar10数据中的模型保存以及测试：
训练以及模型保存代码：
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue Jan 9 21:46:26 2018
@author: lvedng
"""
# In[1]:
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
def read_and_decode(filename): # 读入dog_train.tfrecords
# filename_queue = tf.train.string_input_producer([filename]) # 生成一个queue队列
filename_queue = tf.train.string_input_producer([filename],shuffle=True,num_epochs=100) # 生成一个queue队列
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue) # 返回文件名和文件
features = tf.parse_single_example(serialized_example,
features={
'label': tf.FixedLenFeature([], tf.int64),
'img_raw': tf.FixedLenFeature([], tf.string),
}) # 将image数据和label取出来
img = tf.decode_raw(features['img_raw'], tf.uint8)
img = tf.reshape(img, [32 * 32 * 3]) # reshape为128*128的3通道图片
img = tf.cast(img, tf.float32) * (1. / 255)
label = tf.cast(features['label'], tf.int32) # 在流中抛出label张量
return img, label
def createBatch(filename, batchsize):
images, labels = read_and_decode(filename)
min_after_dequeue = batchsize
capacity = 10* batchsize
image_batch, label_batch = tf.train.shuffle_batch([images, labels],
batch_size=batchsize,
capacity=capacity,
min_after_dequeue=min_after_dequeue
)
label_batch = tf.one_hot(label_batch, depth=10)
return image_batch, label_batch
# 实现Batch Normalization
def bn_layer(x,name, is_training, moving_decay=0.9, eps=1e-5):
# 获取输入维度并判断是否匹配卷积层(4)或者全连接层(2)
shape = x.shape
assert len(shape) in [2, 4]
param_shape = shape[-1]
with tf.variable_scope(name):
# 声明BN中唯一需要学习的两个参数，y=gamma*x+beta
gamma = tf.get_variable("gamma", param_shape, initializer=tf.constant_initializer(1))
beta = tf.get_variable("beta", param_shape, initializer=tf.constant_initializer(0))
# 计算当前整个batch的均值与方差
axes = list(range(len(shape) - 1)) # 特征的通道数
batch_mean, batch_var = tf.nn.moments(x, axes, name='moments')
# 采用滑动平均更新均值与方差
ema = tf.train.ExponentialMovingAverage(moving_decay)
def mean_var_with_update():
ema_apply_op = ema.apply([batch_mean, batch_var])
with tf.control_dependencies([ema_apply_op]):
return tf.identity(batch_mean), tf.identity(batch_var)
# 训练时，更新均值与方差，测试时使用之前最后一次保存的均值与方差
mean, var = tf.cond(tf.equal(is_training, True), mean_var_with_update,
lambda: (ema.average(batch_mean), ema.average(batch_var)))
# 最后执行batch normalization
return tf.nn.batch_normalization(x, mean, var, beta, gamma, eps)
# 初始化权值
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1) # 生成一个截断的正态分布
return tf.Variable(initial)
# 初始化偏置
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def dwconv(name,input, filter_size, in_filters, mutil_filters, strides):
regularizer = tf.contrib.layers.l2_regularizer(scale = 0.0001)
n = filter_size * filter_size * mutil_filters
with tf.variable_scope(name):
weights = tf.get_variable('conv_weights', shape=[filter_size, filter_size, in_filters, mutil_filters],
initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0 / n)),
regularizer=regularizer)
# f2 = tf.Variable(tf.truncated_normal([filter_size, filter_size, in_filters, mutil_filters], stddev=0.1))
return tf.nn.depthwise_conv2d(input, weights, strides, padding="SAME", rate=None, name=None, data_format=None)
# 池化层
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# 2D卷积
def _conv(name, x, filter_size, in_fliters, out_filters, strides):
b = bias_variable([out_filters])
regularizer = tf.contrib.layers.l2_regularizer(scale=0.0001)
n = filter_size * filter_size * out_filters
with tf.variable_scope(name):
# 获取或新建卷积核，正态随机初始化
weights = tf.get_variable('convws',
shape=[filter_size, filter_size, in_fliters, out_filters],
initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0 / n)),
regularizer=regularizer)
# kernel = tf.Variable(tf.truncated_normal([filter_size, filter_size, in_fliters, out_filters], stddev=0.1))
out=tf.nn.conv2d(x, weights, strides, padding='SAME') + b
# 计算卷积
return out
# 把步长值转换成tf.nn.conv2d需要的步长数组
def stride_arr(stride):
return [1, stride, stride, 1]
def output(x_tensor, num_outputs):
#全连接层
x_shape = x_tensor.get_shape().as_list()
regularizer = tf.contrib.layers.l2_regularizer(scale = 0.0001)
weights = tf.get_variable('weight', shape = [x_shape[1], num_outputs],
initializer = tf.uniform_unit_scaling_initializer(factor = 1.0),
regularizer = regularizer)
bias = tf.Variable(tf.zeros([num_outputs]))
out = tf.add(tf.matmul(x_tensor, weights), bias)
return out
def visualize_feature_map_sum(feature_batch):
'''
将每张子图进行相加
:param feature_batch:
:return:
'''
feature_map = feature_batch
feature_map_combination = []
# 取出 featurn map 的数量
num_pic = feature_map.shape[2]
# 将 每一层卷积的特征图，拼接层 5 × 5
for i in range(0, num_pic):
feature_map_split = feature_map[:, :, i]
feature_map_combination.append(feature_map_split)
# 按照特征图 进行 叠加代码
feature_map_sum = sum(one for one in feature_map_combination)
plt.imshow(feature_map_sum)
plt.show()
def visualize_feature_map(feature_batch):
'''
创建特征子图，创建叠加后的特征图
:param feature_batch: 一个卷积层所有特征图
:return:
'''
feature_map = feature_batch
feature_map_combination = []
plt.figure(figsize=(8, 8))
# 取出 featurn map 的数量，因为特征图数量很多，这里直接手动指定了。
#num_pic = feature_map.shape[2]
row, col = get_row_col(64)
# 将 每一层卷积的特征图，拼接层 5 × 5
for i in range(0, 64):
feature_map_split = feature_map[:, :, i]
feature_map_combination.append(feature_map_split)
plt.subplot(row, col, i+1)
plt.imshow(feature_map_split)
plt.axis('off')
def get_row_col(num_pic):
'''
计算行列的值
:param num_pic: 特征图的数量
:return:
'''
squr = num_pic ** 0.5
row = round(squr)
col = row + 1 if squr - row > 0 else row
return row, col
#plt.savefig('./mao_feature/feature_map2.png') # 保存图像到本地
plt.show()
# 命名空间
lamda = 0.004 # 正则化系数，这里添加了正则化但是没有使用
# 定义两个placeholder
x = tf.placeholder(tf.float32, [None, 32 * 32 * 3])
y = tf.placeholder(tf.float32, [None, 10])
# 改变x的格式转为4D的向量[batch, in_height, in_width, in_channels]`
x_image = tf.reshape(x, [-1, 32, 32, 3])
is_training = tf.placeholder(tf.bool)
learnrate = tf.placeholder(tf.float32)
conv1 = _conv("conv1", x_image, 3, 3, 32, stride_arr(2)) # 16
conv1 = bn_layer(conv1, "BN1",is_training)
conv1 = tf.nn.relu(conv1)
conv2 = dwconv('dw2',conv1, 3, 32, 1, stride_arr(1))
conv2 = bn_layer(conv2, "BN2_1", is_training)
conv2 = tf.nn.relu(conv2)
conv2 = _conv("conv2", conv2, 1, 32, 64, stride_arr(1))
conv2 = bn_layer(conv2, "BN2_2", is_training)
conv2 = tf.nn.relu(conv2)
conv3 = dwconv('dw3',conv2, 3, 64, 1, stride_arr(2)) # 8
conv3 = bn_layer(conv3, "BN3_1", is_training)
conv3 = tf.nn.relu(conv3)
conv3 = _conv("conv3", conv3, 1, 64, 128, stride_arr(1))
conv3 = bn_layer(conv3, "BN3_2", is_training)
conv3 = tf.nn.relu(conv3)
conv4 = dwconv('dw4',conv3, 3, 128, 1, stride_arr(1))
conv4 = bn_layer(conv4, "BN4_1", is_training)
conv4 = tf.nn.relu(conv4)
conv4 = _conv("conv4", conv4, 1, 128, 128, stride_arr(1))
conv4 = bn_layer(conv4, "BN4_2", is_training)
conv4 = tf.nn.relu(conv4)
conv5 = dwconv('dw5',conv4, 3, 128, 1, stride_arr(2)) # 4
conv5 = bn_layer(conv5, "BN5_1", is_training)
conv5 = tf.nn.relu(conv5)
conv5 = _conv("conv5", conv5, 1, 128, 256, stride_arr(1))
conv5 = bn_layer(conv5, "BN5_2", is_training)
conv5 = tf.nn.relu(conv5)
conv6 = dwconv('dw6',conv5, 3, 256, 1, stride_arr(1))
conv6 = bn_layer(conv6, "BN6_1", is_training)
conv6 = tf.nn.relu(conv6)
conv6 = _conv("conv6", conv6, 1, 256, 256, stride_arr(1))
conv6 = bn_layer(conv6, "BN6_2", is_training)
conv6 = tf.nn.relu(conv6)
conv7 = dwconv('dw7',conv6, 3, 256, 1, stride_arr(2)) # 2
conv7 = bn_layer(conv7, "BN7_1", is_training)
conv7 = tf.nn.relu(conv7)
conv7 = _conv("conv7", conv7, 1, 256, 512, stride_arr(1))
conv7 = bn_layer(conv7, "BN7_2", is_training)
conv7 = tf.nn.relu(conv7)
#10个重复的层
conv8 = dwconv('dw8_1',conv7, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_1", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_1", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_2", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_2',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_3", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_2", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_4", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_3',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_5", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_3", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_6", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_4',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_7", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_4", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_8", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_5',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_9", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("8_5", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_10", is_training)
conv8 = tf.nn.relu(conv8)
conv9 = dwconv('dw9',conv8, 3, 512, 1, stride_arr(2)) # 1
conv9 = bn_layer(conv9, "BN9_1", is_training)
conv9 = tf.nn.relu(conv9)
conv9 = _conv("conv9", conv9, 1, 512, 1024, stride_arr(1))
conv9 = bn_layer(conv9, "BN9_2", is_training)
conv9 = tf.nn.relu(conv9)
conv10 = dwconv('dw10',conv9, 3, 1024, 1, stride_arr(1)) # 1*1
conv10 = bn_layer(conv10, "BN10_1", is_training)
conv10 = tf.nn.relu(conv10)
conv10 = _conv("conv10", conv10, 1, 1024, 1024, stride_arr(1))
conv10 = bn_layer(conv10,"BN10_2", is_training)
conv10 = tf.nn.relu(conv10)
conv10=tf.reduce_mean(conv10,[1,2])
log = output(conv10,10)
prediction = tf.nn.softmax(log)
# 交叉熵损失
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=log))
# 总损失
zloss = cross_entropy
# 用AdamOptimizer优化器训练
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_step = tf.train.RMSPropOptimizer(learnrate).minimize(zloss)
# 计算准确率
labely=tf.argmax(y, 1)
labelp=tf.argmax(prediction, 1)
correct_prediction = tf.equal(labely, labelp)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) # tf.cast将数据转换成指定类型
# var_list = tf.trainable_variables()
#
# g_list = tf.global_variables()
#
# bn_moving_vars = [g for g in g_list if 'moving_mean' in g.name]
#
# bn_moving_vars += [g for g in g_list if 'moving_variance' in g.name]
#
# var_list += bn_moving_vars
# saver = tf.train.Saver(var_list=var_list, max_to_keep=1)
saver = tf.train.Saver(tf.global_variables())
tfrecord_path = 'train_shuffle.tfrecords'
test_path = 'test_shuffle.tfrecords'
image_batch, label_batch = createBatch(filename=tfrecord_path, batchsize=128)
test_batch, testlabel_batch = createBatch(filename=test_path, batchsize=128)
train_acc = []
train_loss = []
test_acc = []
test_loss = []
with tf.Session() as sess:
sess.run(tf.group(tf.global_variables_initializer(),tf.local_variables_initializer()))
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
rate = 0.001
for i in range(100):
if i<5000:
rate=0.1
elif i>=5000 and i<7000:
rate=0.01
else:
rate=0.001
# 训练模型
img_xs, label_xs = sess.run([image_batch, label_batch])
_, trainacc, loss1 = sess.run([train_step, accuracy, cross_entropy],
feed_dict={is_training: True, x: img_xs, y: label_xs, learnrate: rate,
})
# _, trainacc, loss1,labely1,labelp1 = sess.run([train_step, accuracy, cross_entropy,labely,labelp],
# feed_dict={is_training: True, x: img_xs, y: label_xs, learnrate: rate,
# keep_prob: 0.5})
# print("输出数据标签的类型以及大小")
# print("实际的标签")
# print(labely1)
# print("预测的标签")
# print(labelp1)
print("Itertrainbatch " + str(i) + ", train accuracy= " + str(trainacc) + ",loss=" + str(loss1))
if i != 0:
if i % 10== 0:
train_acc.append(trainacc)
train_loss.append(loss1)
img_txs, label_txs = sess.run([test_batch, testlabel_batch])
conv2s,accur, loss2 = sess.run([conv2,accuracy, cross_entropy],
feed_dict={is_training: False, x: img_txs, y: label_txs, learnrate: rate,
})
# 卷积1
# test_image = np.reshape(conv2s, (128, 16, 16, 64))
# 查看 每个 特征 子图
# visualize_feature_map(test_image[0])
# 查看 叠加后的 特征图
# visualize_feature_map_sum(test_image[0])
# fig2, ax2 = plt.subplots(nrows=1, ncols=64, figsize=(64, 1))
# for i in range(64):
# ax2[i].imshow(np.reshape(test_image[0], (64, 1, 16, 16))[i][0])
# test_imagesum=test_image[0]
#
# plt.title('Conv2 32x14x14')
#
# plt.show()
test_acc.append(accur)
test_loss.append(loss2)
# print('11111111111111111111111111111111111111111111111111111111111111111111111')
# print("输出测试数据标签的类型以及大小")
# print("实际测试的标签")
# print(labely2)
# print("预测测试的标签")
# print(labelp2)
print("Iter " + str(i) + ", Test accuracy= " + str(accur) + ",loss=" + str(loss2))
#
# print('11111111111111111111111111111111111111111111111111111111111111111111111')
coord.request_stop()
coord.join(threads)
# 保存模型
saver.save(sess, 'net/my_net.ckpt')
fig = plt.figure()
ax1 = fig.add_subplot(111)
ax1.plot(test_acc, 'r')
ax1.plot(train_acc, 'b')
ax1.set_ylabel('accuracy')
# ax1.set_title("Double Y axis")
ax2 = ax1.twinx() # this is the important function
ax2.plot(test_loss, 'r')
ax2.plot(train_loss, 'b')
ax2.set_ylabel('loss')
plt.show()
测试代码：
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue Jan 9 21:46:26 2018
@author: lvedng
"""
# In[1]:
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image #注意Image,后面会用到
def read_and_decode(filename): # 读入dog_train.tfrecords
# filename_queue = tf.train.string_input_producer([filename]) # 生成一个queue队列
filename_queue = tf.train.string_input_producer([filename],shuffle=True,num_epochs=100) # 生成一个queue队列
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue) # 返回文件名和文件
features = tf.parse_single_example(serialized_example,
features={
'label': tf.FixedLenFeature([], tf.int64),
'img_raw': tf.FixedLenFeature([], tf.string),
}) # 将image数据和label取出来
img = tf.decode_raw(features['img_raw'], tf.uint8)
img = tf.reshape(img, [32 * 32 * 3]) # reshape为128*128的3通道图片
img = tf.cast(img, tf.float32) * (1. / 255)
label = tf.cast(features['label'], tf.int32) # 在流中抛出label张量
return img, label
def createBatch(filename, batchsize):
images, labels = read_and_decode(filename)
min_after_dequeue = batchsize
capacity = 10* batchsize
image_batch, label_batch = tf.train.shuffle_batch([images, labels],
batch_size=batchsize,
capacity=capacity,
min_after_dequeue=min_after_dequeue
)
label_batch = tf.one_hot(label_batch, depth=10)
return image_batch, label_batch
# 实现Batch Normalization
def bn_layer(x,name, is_training, moving_decay=0.9, eps=1e-5):
# 获取输入维度并判断是否匹配卷积层(4)或者全连接层(2)
shape = x.shape
assert len(shape) in [2, 4]
param_shape = shape[-1]
with tf.variable_scope(name):
# 声明BN中唯一需要学习的两个参数，y=gamma*x+beta
gamma = tf.get_variable("gamma", param_shape, initializer=tf.constant_initializer(1))
beta = tf.get_variable("beta", param_shape, initializer=tf.constant_initializer(0))
# 计算当前整个batch的均值与方差
axes = list(range(len(shape) - 1)) # 特征的通道数
batch_mean, batch_var = tf.nn.moments(x, axes, name='moments')
# 采用滑动平均更新均值与方差
ema = tf.train.ExponentialMovingAverage(moving_decay)
def mean_var_with_update():
ema_apply_op = ema.apply([batch_mean, batch_var])
with tf.control_dependencies([ema_apply_op]):
return tf.identity(batch_mean), tf.identity(batch_var)
# 训练时，更新均值与方差，测试时使用之前最后一次保存的均值与方差
mean, var = tf.cond(tf.equal(is_training, True), mean_var_with_update,
lambda: (ema.average(batch_mean), ema.average(batch_var)))
# 最后执行batch normalization
return tf.nn.batch_normalization(x, mean, var, beta, gamma, eps)
# 初始化权值
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1) # 生成一个截断的正态分布
return tf.Variable(initial)
# 初始化偏置
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def dwconv(name,input, filter_size, in_filters, mutil_filters, strides):
regularizer = tf.contrib.layers.l2_regularizer(scale = 0.0001)
n = filter_size * filter_size * mutil_filters
with tf.variable_scope(name):
weights = tf.get_variable('conv_weights', shape=[filter_size, filter_size, in_filters, mutil_filters],
initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0 / n)),
regularizer=regularizer)
# f2 = tf.Variable(tf.truncated_normal([filter_size, filter_size, in_filters, mutil_filters], stddev=0.1))
return tf.nn.depthwise_conv2d(input, weights, strides, padding="SAME", rate=None, name=None, data_format=None)
# 池化层
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# 2D卷积
def _conv(name, x, filter_size, in_fliters, out_filters, strides):
b = bias_variable([out_filters])
regularizer = tf.contrib.layers.l2_regularizer(scale=0.0001)
n = filter_size * filter_size * out_filters
with tf.variable_scope(name):
# 获取或新建卷积核，正态随机初始化
weights = tf.get_variable('convws',
shape=[filter_size, filter_size, in_fliters, out_filters],
initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0 / n)),
regularizer=regularizer)
# kernel = tf.Variable(tf.truncated_normal([filter_size, filter_size, in_fliters, out_filters], stddev=0.1))
out=tf.nn.conv2d(x, weights, strides, padding='SAME') + b
# 计算卷积
return out
# 把步长值转换成tf.nn.conv2d需要的步长数组
def stride_arr(stride):
return [1, stride, stride, 1]
def output(x_tensor, num_outputs):
#全连接层
x_shape = x_tensor.get_shape().as_list()
regularizer = tf.contrib.layers.l2_regularizer(scale = 0.0001)
weights = tf.get_variable('weight', shape = [x_shape[1], num_outputs],
initializer = tf.uniform_unit_scaling_initializer(factor = 1.0),
regularizer = regularizer)
bias = tf.Variable(tf.zeros([num_outputs]))
out = tf.add(tf.matmul(x_tensor, weights), bias)
return out
def get_one_image(img_dir):
image = Image.open(img_dir)
plt.imshow(image)
image = image.resize([32,32])
image_arr = np.array(image)
return image_arr
# 命名空间
lamda = 0.004 # 正则化系数，这里添加了正则化但是没有使用
# 定义两个placeholder
x = tf.placeholder(tf.float32, [1, 32 * 32 * 3])
y = tf.placeholder(tf.float32, [1, 10])
# 改变x的格式转为4D的向量[batch, in_height, in_width, in_channels]`
x_image = tf.reshape(x, [1, 32, 32, 3])
is_training = tf.placeholder(tf.bool)
learnrate = tf.placeholder(tf.float32)
conv1 = _conv("conv1", x_image, 3, 3, 32, stride_arr(2)) # 16
conv1 = bn_layer(conv1, "BN1",is_training)
conv1 = tf.nn.relu(conv1)
conv2 = dwconv('dw2',conv1, 3, 32, 1, stride_arr(1))
conv2 = bn_layer(conv2, "BN2_1", is_training)
conv2 = tf.nn.relu(conv2)
conv2 = _conv("conv2", conv2, 1, 32, 64, stride_arr(1))
conv2 = bn_layer(conv2, "BN2_2", is_training)
conv2 = tf.nn.relu(conv2)
conv3 = dwconv('dw3',conv2, 3, 64, 1, stride_arr(2)) # 8
conv3 = bn_layer(conv3, "BN3_1", is_training)
conv3 = tf.nn.relu(conv3)
conv3 = _conv("conv3", conv3, 1, 64, 128, stride_arr(1))
conv3 = bn_layer(conv3, "BN3_2", is_training)
conv3 = tf.nn.relu(conv3)
conv4 = dwconv('dw4',conv3, 3, 128, 1, stride_arr(1))
conv4 = bn_layer(conv4, "BN4_1", is_training)
conv4 = tf.nn.relu(conv4)
conv4 = _conv("conv4", conv4, 1, 128, 128, stride_arr(1))
conv4 = bn_layer(conv4, "BN4_2", is_training)
conv4 = tf.nn.relu(conv4)
conv5 = dwconv('dw5',conv4, 3, 128, 1, stride_arr(2)) # 4
conv5 = bn_layer(conv5, "BN5_1", is_training)
conv5 = tf.nn.relu(conv5)
conv5 = _conv("conv5", conv5, 1, 128, 256, stride_arr(1))
conv5 = bn_layer(conv5, "BN5_2", is_training)
conv5 = tf.nn.relu(conv5)
conv6 = dwconv('dw6',conv5, 3, 256, 1, stride_arr(1))
conv6 = bn_layer(conv6, "BN6_1", is_training)
conv6 = tf.nn.relu(conv6)
conv6 = _conv("conv6", conv6, 1, 256, 256, stride_arr(1))
conv6 = bn_layer(conv6, "BN6_2", is_training)
conv6 = tf.nn.relu(conv6)
conv7 = dwconv('dw7',conv6, 3, 256, 1, stride_arr(2)) # 2
conv7 = bn_layer(conv7, "BN7_1", is_training)
conv7 = tf.nn.relu(conv7)
conv7 = _conv("conv7", conv7, 1, 256, 512, stride_arr(1))
conv7 = bn_layer(conv7, "BN7_2", is_training)
conv7 = tf.nn.relu(conv7)
#10个重复的层
conv8 = dwconv('dw8_1',conv7, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_1", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_1", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_2", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_2',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_3", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_2", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_4", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_3',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_5", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_3", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_6", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_4',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_7", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("conv8_4", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_8", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = dwconv('dw8_5',conv8, 3, 512, 1, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_9", is_training)
conv8 = tf.nn.relu(conv8)
conv8 = _conv("8_5", conv8, 1, 512, 512, stride_arr(1))
conv8 = bn_layer(conv8, "BN8_10", is_training)
conv8 = tf.nn.relu(conv8)
conv9 = dwconv('dw9',conv8, 3, 512, 1, stride_arr(2)) # 1
conv9 = bn_layer(conv9, "BN9_1", is_training)
conv9 = tf.nn.relu(conv9)
conv9 = _conv("conv9", conv9, 1, 512, 1024, stride_arr(1))
conv9 = bn_layer(conv9, "BN9_2", is_training)
conv9 = tf.nn.relu(conv9)
conv10 = dwconv('dw10',conv9, 3, 1024, 1, stride_arr(1)) # 1*1
conv10 = bn_layer(conv10, "BN10_1", is_training)
conv10 = tf.nn.relu(conv10)
conv10 = _conv("conv10", conv10, 1, 1024, 1024, stride_arr(1))
conv10 = bn_layer(conv10,"BN10_2", is_training)
conv10 = tf.nn.relu(conv10)
conv10=tf.reduce_mean(conv10,[1,2])
log = output(conv10,10)
# 将8 * 8 * 64 三维向量拉直成一行向量
# h_pool2_flat = tf.reshape(conv10, [-1, 1 * 1 * 512])
# # 第一层全连接
# W_fc1 = weight_variable([1 * 1 * 512, 256])
# b_fc1 = bias_variable([256])
# h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
# # # 对隐藏层使用dropout
#
# keep_prob = tf.placeholder(tf.float32)
# W_fc2 = weight_variable([256, 10])
# b_fc2 = bias_variable([10])
# log = tf.matmul(h_fc1, W_fc2) + b_fc2
prediction = tf.nn.softmax(log)
# w1_loss = lamda * tf.nn.l2_loss(W_fc1) # 对W_fc1使用L2正则化
# w2_loss = lamda * tf.nn.l2_loss(W_fc2) # 对W_fc2使用L2正则化
# 交叉熵损失
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=log))
# 总损失
zloss = cross_entropy
# 用AdamOptimizer优化器训练
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_step = tf.train.RMSPropOptimizer(learnrate).minimize(zloss)
# 计算准确率
labely=tf.argmax(y, 1)
labelp=tf.argmax(prediction, 1)
correct_prediction = tf.equal(labely, labelp)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) # tf.cast将数据转换成指定类型
saver = tf.train.Saver(max_to_keep=1)
image_arr = get_one_image('G:\\data\\test\\00000\\3215.png')
image_arr=image_arr.reshape([1,32 * 32 * 3])
with tf.Session() as sess:
saver.restore(sess, 'net/my_net.ckpt') # 注意此处路径前添加"./"
prediction1,labelp1, = sess.run([prediction,labelp], feed_dict={is_training: False,x: image_arr})
print(prediction1)
print(labelp1)
扫一扫
635
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
