# 全局初始化 tf.global_variables_initializer().run() # 等价于 init = tf.global_variables_initializer() self.sess = tf.Session() self.sess.run(init)
# 优化的分类
optimizer = tf.train.AdamOptimizer()
self.sess = tf.Session() self.cost = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(self.reconstruction, self.x), 2.0)) self.optimizer = optimizer.minimize(self.cost) # 等价于self.optimizer=tf.train.AdamOptimizer().minimize(self.cost)
def partial_fit(self, X):
cost, opt = self.sess.run((self.cost, self.optimizer), feed_dict = {self.x: X, self.scale: self.training_scale }) # 等价于cost,opt = tf.train.Ada,Optimizer().minimize(self.cost).run(feed_dict)
return cost
# 第二种
train_step = tf.train.GradientDescentOptimizer(0.2).minimize(cross_entropy)
train_step.run({x: batch_xs, y_: batch_ys})
# 使用自适应的优化器Adagrad train_step = tf.train.AdagradOptimizer(0.3).minimize(cross_entropy)
# 简单的
self.optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(self.loss)
# 文件保存 self.summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph())
# 编码和解码
def encoder(self,x, weights):
shapes = []
# Encoder Hidden layer with relu activation #1
shapes.append(x.get_shape().as_list())
layer1 = tf.nn.bias_add(tf.nn.conv2d(x, weights['enc_w0'], strides=[1,2,2,1],padding='SAME'),weights['enc_b0'])
layer1 = tf.nn.relu(layer1) # 激活函数
shapes.append(layer1.get_shape().as_list())
layer2 = tf.nn.bias_add(tf.nn.conv2d(layer1, weights['enc_w1'], strides=[1,2,2,1],padding='SAME'),weights['enc_b1'])
layer2 = tf.nn.relu(layer2)
shapes.append(layer2.get_shape().as_list())
layer3 = tf.nn.bias_add(tf.nn.conv2d(layer2, weights['enc_w2'], strides=[1,2,2,1],padding='SAME'),weights['enc_b2'])
layer3 = tf.nn.relu(layer3)
return layer3, shapes
# Building the decoder
def decoder(self,z, weights, shapes):
# Encoder Hidden layer with relu activation #1
shape_de1 = shapes[2]
layer1 = tf.add(tf.nn.conv2d_transpose(z, weights['dec_w0'], tf.stack([tf.shape(self.x)[0],shape_de1[1],shape_de1[2],shape_de1[3]]),
strides=[1,2,2,1],padding='SAME'),weights['dec_b0'])
layer1 = tf.nn.relu(layer1)
shape_de2 = shapes[1]
layer2 = tf.add(tf.nn.conv2d_transpose(layer1, weights['dec_w1'], tf.stack([tf.shape(self.x)[0],shape_de2[1],shape_de2[2],shape_de2[3]]),
strides=[1,2,2,1],padding='SAME'),weights['dec_b1'])
layer2 = tf.nn.relu(layer2)
shape_de3= shapes[0]
layer3 = tf.add(tf.nn.conv2d_transpose(layer2, weights['dec_w2'], tf.stack([tf.shape(self.x)[0],shape_de3[1],shape_de3[2],shape_de3[3]]),
strides=[1,2,2,1],padding='SAME'),weights['dec_b2'])
layer3 = tf.nn.relu(layer3)
return layer3
# batch选择
def next_batch(data, _index_in_epoch ,batch_size , _epochs_completed):
_num_examples = data.shape[0]
start = _index_in_epoch
_index_in_epoch += batch_size
if _index_in_epoch > _num_examples:
# Finished epoch
_epochs_completed += 1
# Shuffle the data
perm = np.arange(_num_examples) # 创建一个0-np.arange的等差数组
np.random.shuffle(perm) # 随机打乱列表
data = data[perm]
#label = label[perm]
# Start next epoch
start = 0
_index_in_epoch = batch_size
assert batch_size <= _num_examples # 判断
end = _index_in_epoch
return data[start:end], _index_in_epoch, _epochs_completed
如何获得tensorflow各层的参数呢？
def getWeights(self): # 获得隐含层的权重
return self.sess.run(self.weights['w1'])
# 这里的self.weights['w1'] 需要在前面定义过
def print_activations(t):
print(t.op.name, ' ', t.get_shape().as_list())
数据预处理代码：
# 数据零均值，特征方差归一化处理
def standard_scale(X_train, X_test):
preprocessor = prep.StandardScaler().fit(X_train)
X_train = preprocessor.transform(X_train)
X_test = preprocessor.transform(X_test)
return X_train, X_test
batch样本选取
# batch随机选取
def get_random_block_from_data(data, batch_size): start_index = np.random.randint(0, len(data) - batch_size) return data[start_index:(start_index + batch_size)]
# 参数初始化的方法 np.random.normal； np.random.randint； tf.zeros；# tf.random_uniform
def _initialize_weights(self):
all_weights = dict()
all_weights['w1'] = tf.Variable(xavier_init(self.n_input, self.n_hidden))
all_weights['b1'] = tf.Variable(tf.zeros([self.n_hidden], dtype = tf.float32))
all_weights['w2'] = tf.Variable(tf.zeros([self.n_hidden, self.n_input], dtype = tf.float32))
all_weights['b2'] = tf.Variable(tf.zeros([self.n_input], dtype = tf.float32))
return all_weights
hidden = np.random.normal(size = self.weights["b1"]) # 随机噪声
start_index = np.random.randint(0, len(data) - batch_size)
def xavier_init(fan_in, fan_out, constant = 1): # 标准的均匀分布
low = -constant * np.sqrt(6.0 / (fan_in + fan_out))
high = constant * np.sqrt(6.0 / (fan_in + fan_out))
return tf.random_uniform((fan_in, fan_out),
minval = low, maxval = high,
dtype = tf.float32)
W1 = tf.Variable(tf.truncated_normal([in_units, h1_units], stddev=0.1)) # 截断的正太分布，方差为0.1
激活函数：
tf.nn.softplus
tf.nn.relu
tensorflow的卷积
weight2 = variable_with_weight_loss(shape=[5, 5, 64, 64], stddev=5e-2, wl=0.0) # 第二个卷积层权重的定义
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
net = slim.conv2d(inputs, 32, [3, 3], stride=2, scope='Conv2d_1a_3x3') #用slim定义卷积
tensorflow的尺寸转换
x_image = tf.reshape(x, [-1,28,28,1]) # 将1*784的向量转换为28*28的图像
# 卷积层的定义 W_conv1 = weight_variable([5, 5, 1, 32]) b_conv1 = bias_variable([32]) h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) h_pool1 = max_pool_2x2(h_conv1)
tensorflow获取卷积，池化之后数据的维度函数：
shape = h_pool2.get_shape()[1].value # 这样可以获取第2个维度
def print_activations(t):
print(t.op.name, ' ', t.get_shape().as_list())
卷积层命名
# 将 scope内生成的变量自动命名为conv1
with tf.name_scope('conv1') as scope: kernel = tf.Variable(tf.truncated_normal([11, 11, 3, 64], dtype=tf.float32, stddev=1e-1), name='weights') conv = tf.nn.conv2d(images, kernel, [1, 4, 4, 1], padding='SAME') biases = tf.Variable(tf.constant(0.0, shape=[64], dtype=tf.float32), trainable=True, name='biases') bias = tf.nn.bias_add(conv, biases) conv1 = tf.nn.relu(bias, name=scope) print_activations(conv1) # 打印卷积层1的输出结果 parameters += [kernel, biases]
执行计算
config = tf.ConfigProto() config.gpu_options.allocator_type = 'BFC' sess = tf.Session(config=config) # 可以这么替换 sess = tf.Session()
Block('block1', bottleneck, [(256, 64, 1)] * 2 + [(256, 64, 2)])
# 其中block1是名称，bottleneck是ResNet的残差学习单元，（256,64,1）：弟三层的通道数深度是256，前两层的通道数是64，中间那层的步长是3；结构为：【(1*1/S1,64),(3*3/s2,64),(1*1/S1,256)】，
# 在这个Block中，一共有3个bottleneck残差学习单元，除了最后一个步长由3变为2，其余都一样。
一般我们用append得到的是一个list型数据，此时可以用np.array转换为一个三维数组
I = []
Label = []
for i in range(img.shape[2]): # img.shape[2] = 38 个类
for j in range(img.shape[1]): # img.shape[1] = 64 个样本/类
temp = np.reshape(img[:,j,i],[42,48]) # 将向量转换为矩阵
Label.append(i) # list
I.append(temp)
用I.append得到的I
用np.array转换成的矩阵
然后使用Img = np.transpose(I,[0,2,1])得到的数据
然后使用Img = np.expand_dims(Img,3)
perm = np.arange(_num_examples) # 创建一个0-np.arange的等差数组 [0,1,2,3,..10] np.random.shuffle(perm) # 随机打乱列表 [0,2,6,7,..]
# 模型的保存
def save_model(self):
save_path = self.saver.save(self.sess,self.model_path)
print ("model saved in file: %s" % save_path)
# 模型的恢复
def restore(self):
self.saver.restore(self.sess, self.restore_path)
print ("model restored")