def loss(self,output, l_out):
with tf.variable_scope('loss') as scope:
nonzero_bool = tf.not_equal(output, tf.contant(0, tf.float32))
nonzero_mat = tf.cast(nonzero_bool, tf.float32)
l_out_nonzero = tf.multiply(l_out, nonzero_mat)
cross_entropy = tf.square(tf.subtract(l_out_nonzero, output))
cost = tf.reduce_mean(cross_entropy, name=scope.name)
return cost
def batch_norm_template(inputs, is_training, scope, moments_dims, bn_decay): ''' Args: inputs: Tensor 4D, BHWC input maps is_training : boolean tf.Variable , true indicates that training phase bn_decay : float or float tensor variable, controling the moving average weight scope: string, varible scope moments_dims: a list if ints, indicating dimensions for moments calculation Return: normed: batch normalized maps ''' with tf.variable_scope(scope) as sc: num_channels = inputs.get_shape()[-1].value beta = tf.Variable(tf.constant(0.0, shape=[num_channels]), name = 'beta', trainable = True ) gamma = tf.Variable(tf.contant(1.0, shape=[num_channels]), name='gamma', trainable = True ) batch_mean, batch_var = tf.nn.moments(inputs, moments_dims, name='moments') decay = bn_decay if bn_decay is not None else 0.9 ema = tf.train.ExponentialMovingAverage(decay=decay) #operator that maintains the moving averages of variables ema_apply_op = tf.cond(is_training, lambda:ema.apply([batch_mean, batch_var]), lambda:tf.no_op()) def mean_var_with_update(): with tf.control_dependencies([ema_apply_op]): return tf.identity(batch_mean), tf.identity(batch_var) #ema.average returns the Variable holding the average of var mean, var = tf.cond(is_training, mean_var_with_update, lambda: (ema.average(batch_mean), ema.average(batch_var))) normed = tf.nn.batch_normalization(inputs, mean, var, beta, gamma, 1e-3) return normed
x = tf.map_fn(lambda img: tf.image.random_flip_left_right(img), x)
return x
XX = tf.cond(is_training, lambda: aug_image(X), lambda: X)
logits = cnn(XX, is_training=is_training)
clf_loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y))
reg_loss_list = []
for v in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
if re.search('weights', v.name):
reg_loss_list.append(l2(v))
logging.info('Apply {} for {}'.format(l2.__name__, v.name))
reg_loss = tf.add_n(reg_loss_list) if reg_loss_list else tf.contant(
0.0, dtype=tf.float32)
loss = clf_loss + reg_loss
coords_check = tf.get_collection('70f92c137c01d89c6477c5ef22411bfe')
coords_w = coords_check[0][0]
coords_h = coords_check[0][1]
grid_weights = []
for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
if re.search('GRID', v.name):
grid_weights.append(v)
convbn_weights = [
v for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
if v not in grid_weights
]
def inference(images):
parameters = []
# the first convolution layer
# name_scope: like C++'s namespace
with tf.name_scope('conv1')as scope:
# convolution kernel's size = 11*11 and this color passageway has 3 and have 64 convolution kernel
kernel = tf.Variable(tf.truncated_normal([11,11,3,64],dtype=tf.float32,stddev=1e-1),name='weights')
#conv = [batch_size,heigth/4 up,width/4 up,convolution_num]
# https ://www.cnblogs.com/lovepysics/p/7220111.html hint: comput the convolution layer
conv = tf.nn.conv2d(images,kernel,[1,4,4,1],padding='SAME')
# form the tensor which shape is 1*64=[1,1,1,1,....,1]has 64 1 elements per line
biases = tf.Variable(tf.constant(0.0,shape=[64],dtype=tf.float32),trainable=True,name='bias')
# output shape = conv shape
bias = tf.nn.bias_add(conv,biases)
# output shape = bias shape
conv1 = tf.nn.relu(bias,name=scope)
print_activations(conv1)
parameters+=[kernel,biases]
# lrn1 shape = conv1 shape
# have some problems
# depth_radius =4, bias =1.0,alpha =0.001/9,beta=0.75 alexnet's rendementation
lrn1 = tf.nn.lrn(conv1,4,bias=1.0,alpha=0.001/9,beta=0.75,name='lrn1')
# pool1 shape = [batch_size,(height-3)/2+1,(width-3)/2+1,convolution-kernel_size]
pool1 = tf.nn.max_pool(lrn,ksize=[1,3,3,1],strides=[1,2,2,1],padding='VALID',name='pool1')
print_activations(pool1)
with tf.name_scope('conv2')as scope:
kernel1 = tf.Variable(tf.truncated_normal([5,5,64,192],dtype=tf.float32,stddev=1e-1),name='weights')
conv = tf.nn.conv2d(pool1,kernel,[1,1,1,1],padding='SAME')
bias = tf.Variable(tf.contant(0.0,shape=[192],dtype=tf.float32),trainable=True,name='biases')
bias = tf.nn.bias_add(conv,biases)
conv2 = tf.nn.relu(bias,name=scope)
parameters += [kernel,biases]
print_activations(conv2)
lrn2 = tf.nn.lrn(conv2,4,bias=1.0,alpha=0.001/9,beta=0.75,name='lrn2')
pool2 = tf.nn.max_pool(lrn2,ksize=[1,3,3,1],strides=[1,2,2,1],padding='VAILD',name='pool2')
print_activations(pool2)
with tf.name_scope('conv3')as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,192,384],dtype=tf.float32,stddev=1e-1),name='SAME')
conv = tf.nn.conv2d(pool2,kernel,[1,1,1,1],padding="SAME")
biases = tf.Variable(tf/constant(0.0,shape=[384],dtype=tf.float32),trainable=True,name='biases')
bias = tf.nn.bias_add(conv,biases)
conv3 = tf.nn.relu(bias,name=scope)
paremeters += [kernel,biases]
with tf.name_scope('conv4')as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,384,256],dtype=tf.float32,stddev=1e-1),name='weights')
conv = tf.nn.conv2d(conv3,kernel,[1,1,1,1],padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32),trainable=True,name='biases')
bias = tf.nn.bias_add(conv,biases)
conv4 = tf.nn.relu(bias,name=scope)
parameters += [kernel,biases]
print_activations(conv4)
with tf.name_scope('conv5')as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,256,256],dtype=tf.float32,stddev=1e-1),name='weights')
conv = tf.nn.conv2d(conv4,kernel,[1,1,1,1],padding="SAME")
biases = tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32),trainable=True,name='biases')
bias = tf.nn.bias_add(conv,biases)
conv5 = tf.nn.relu(bias,name=scope)
parameters +=[kernel,biases]
print_activations(conv5)
pool5 = tf.nn.max_pool(conv5,ksize=[1,3,3,1],strides=[1,2,2,1],padding="VALID",name='pool5')
print_activations(pool5)
return pool5,parameters
import tensorflow as tf
hello = tf.contant('hello')
sess = tf.session()
prinnt(sess.run(hello))
import cv2
print('openCV')
# 不同图片的压缩比
import cv2
# read_type: 1 彩色图片
img = cv2.imread(file_name, read_type)
# IMWRITE_JPG_QUALIFY 图片质量，
# numbers 0~100
# jpg的压缩质量可以控制，即有损压缩，无法设置透明度
# png是无损压缩，还可以设置透明度
cv2.imwrite(file_name, img, [cv2.IMWRITE_JPG_QUALIFY, numbers])
# numbers: 0~9 值越小，压缩效果越差
cv2.imwrite(file_name, img, [cv2.IMWRITE_PNG_COMPRESSION, numbers])
## 像素处理
img = cv2.imread(file_name, 1)
# 100行100列对应的像素
(b,g,r) = img[100,100]
# tf基本信息
data1 = tf.constant(2.5)
data2 = tf.Variable(10, name='var')
data3 = tf.constant(2.5, dtype=tf.int32)
print(data1)
[1,0,0]
[0,1,0]]
'''
#scatter 有目的性的根据坐标对值进行更新
'''
tf.scatter_nd(
indices, 根据坐标改变对应底板上的值
updates,
shape shape给定底板，一维二维多维，所有值默认为0
'''
indices = tf.constant([[4], [3], [1], [7]])
updates = tf.constant([9, 10, 11, 12])
#底板上下标4的位置更新为9，下标3的位置更新为10，下标1的位置更新为11，下标7的位置更新为12
shape = tf.contant([8]) #底板是一维，值全为0、长度为8的tensor
tf.scatter_nd(indices, updates, shape)
#返回[0,11,0,10,9,0,0,12]
#tf.scatter_nd只能更新底板为0的值
#scatter使用复杂，详情看课时47、48
#meshgrid 应用在48课时，不好笔述，记得复习
'''
给定x和y的范围，生成[x1,y1],[x2,y2],[x3,y3]...等范围内的坐标
'''
y = tf.linspace(-2., 2., 5) #生成-2.到2.的五个值,均匀分布
x = tf.linspace(-2., 2., 5)
point_x, point_y = tf.meshgrid(x, y)
point_x.shape #[5,5],共25个值，每一列上的值都相等
point_y.shape #[5,5],共25个值，每一行上的值都相等
def start(x_in, y_in, signal2model):
E = np.random.uniform(-np.sqrt(1. / signal2model.signal_dim), np.sqrt(1. / signal2model.signal_dim),
(signal2model.hidden_dim, signal2model.signal_dim))
U = np.random.uniform(-np.sqrt(1. / signal2model.hidden_dim), np.sqrt(1. / signal2model.hidden_dim),
(signal2model.hidden_dim, signal2model.hidden_dim))
W = np.random.uniform(-np.sqrt(1. / signal2model.hidden_dim), np.sqrt(1. / signal2model.hidden_dim),
(signal2model.hidden_dim, signal2model.hidden_dim))
V = np.random.uniform(-np.sqrt(1. / signal2model.hidden_dim), np.sqrt(1. / signal2model.hidden_dim),
(signal2model.signal_dim, signal2model.hidden_dim))
b = np.zeros((signal2model.hidden_dim))
c = np.zeros((signal2model.signal_dim))
E = tf.Variable(E, name="E")
U = tf.Variable(U, name="U")
W = tf.Variable(W, name="W")
b = tf.Variable(b, name="b")
c = tf.Variable(c, name="c")
x = tf.contant(x_in)
y = tf.contant(y_in)
var = tf.global_variables_initializer()
coversion_ones = tf.ones((signal2model.mini_batch_size, 1))
def GRU(i, U, W, b, x_0, s_previous):
U_copy, W_copy = U, W
b1 = tf.matmul(coversion_ones, b[i * 3, :])
b2 = tf.matmul(coversion_ones, b[i * 3 + 1, :])
b3 = tf.matmul(coversion_ones, b[i * 3 + 2, :])
z = tf.sigmoid(tf.add(tf.add(tf.matmul(U[i * 3 + 0], x_0), tf.matmul(W[i * 3 + 0],s_previous)), b1))
r = tf.sigmoid(tf.add(tf.add(tf.matmul(U[i * 3 + 1], x_0), tf.matmul(W[i * 3 + 1], s_previous)), b2))
s_candidate = tf.tanh(tf.add(tf.add(tf.matmul(U_copy[i * 3 + 2], x_0),
tf.matmul(W[i * 3 + 2], tf.matmul(s_previous, r)), b3)))
return tf.add(tf.matmul(tf.ones_like(z) - z, s_candidate), tf.matmul(z, s_previous))
def forward_prop_step(x_t, s_prev):
# Embedding layer
x_0 = E[:, x_t[:, 0]]
s = tf.zeros_like(s_prev)
# GRU BLOCK 1 [1 vs 1] #############
# GRU Layer 1
# print(x_e0)
# print(s_prev)
s[0] = GRU(0, U, W, b, x_0, s_prev[0])
s[1] = GRU(1, U, W, b, s[0], s_prev[1])
s[2] = GRU(2, U, W, b, s[1], s_prev[2])
# Final output calculation
# FIRST DIMENSION:
o_t =tf.sparse_softmax()
T.stack(T.nnet.softmax((V[0].dot(s_[:, 0, 1]) + conversion_ones.dot(c[0])))[0],
T.nnet.softmax((V[1].dot(s_[:, 1, 1]) + conversion_ones.dot(c[1])))[0],
T.nnet.softmax((V[2].dot(s_[:, 2, 1]) + conversion_ones.dot(c[2])))[0],
T.nnet.softmax((V[3].dot(s_[:, 3, 1]) + conversion_ones.dot(c[3])))[0])
return [o_t, s_]
import tensorflow as tf ### 创建图开始 ### # initial a constant variable # 1 x 2 创建一个矩阵 op matrix1 = tf.contant([[3., 3.]]) # 创建另一个2 x 1矩阵 op matrix2 = tf.constant([[2.], [4.]]) # 创建一个矩阵乘法 op product = tf.matmul(matrix1, matrix2) ### 创建图结束 ### ### 通过创建会话session来运行图
