tensorflow 按维度相加_TensorFlow一些函数（1）

最新推荐文章于 2023-04-04 11:10:31 发布

weixin_39946534

最新推荐文章于 2023-04-04 11:10:31 发布

阅读量990

点赞数

文章标签： tensorflow 按维度相加

记录几个tensorflow函数用法

tf.unique_with_counts(x,out_idx=tf.dtypes.int32,name=None)

x表示一个一维张量，此函数用来统计x中不同元素的信息，可以统计标签信息

import tensorflow as tf
import numpy as np
x = tf.placeholder(dtype=tf.int32,shape=(10,))
y,idx,count = tf.unique_with_counts(x)

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    x_v = np.random.choice(np.arange(0,10),size=10,replace=True)
    print('x:',x_v)
    y_v,idx_v,count_v = sess.run([y,idx,count],feed_dict={x:x_v})
    print('y:',y_v)
    print('idx:',idx_v)
    print('count',count_v)

output:
x: [5 0 6 9 6 1 9 9 1 5]
y: [5 0 6 9 1]# 元素集合
idx: [0 1 2 3 2 4 3 3 4 0]# x中元素在y中对应的索引
count [2 1 2 3 2]# 每种元素的个数

tf.scatter_sub(ref,indices,updates,use_locking=False,name=None)

ref表示张量（大于零维），将ref中的值按照indices，用updates更新，ref按照第一个维度进行索引，对应的还有scatter_add,scatter_div,scatter_mul,scatter_max,scatter_min

其中max和min将updates和ref对应的值进行比较，返回较大或者较小的那个

indices和updates第一个维度必须相同，如果indices表示标签，那么该函数可以按标签更新，维护一个类似类中心的变量

import tensorflow as tf
import numpy as np
ref = tf.Variable([1,3,5,7,9,2,4,6,8,10],dtype = tf.int32)
index = tf.constant([4, 3, 1, 7],dtype = tf.int32)
updates = tf.constant([9, 7, 3, 6],dtype = tf.int32)
ref_sub = tf.scatter_sub(ref,index,updates)

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    ref_sub_v = sess.run(ref_sub)
    print(ref_sub_v)
output:[ 1  0  5  0  0  2  4  0  8 10]

对应索引的原始值减去更新值

tf.scatter_nd(indices,updates,shape,name=None)

scatter_nd用零初始化一个shape维度的张量，然后用indices作为索引，用updates的值初始化新创建的全零张量，这里应该要求indices的索引值不能超过shape的第一个维度（不然索引找不到）

tf.scatter_nd(indices, values, shape) 和 tf.tensor_scatter_add(tf.zeros(shape, values.dtype), indices, values)表示的含义相同

import tensorflow as tf
import numpy as np
indices = tf.constant([[4], [3], [1], [7]])
updates = tf.constant([9, 10, 11, 12])
shape = tf.constant([8])
scatter = tf.scatter_nd(indices, updates, shape)
with tf.Session() as sess:
    print(sess.run(scatter))
output:[ 0 11  0 10  9  0  0 12]

借用官网的图https://www.tensorflow.org/versions/r1.14/api_docs/python/tf/scatter_nd?hl=zh-cn

tf.scatter_nd_add(ref,indices,updates,use_locking=False,name=None)

升级版的scatter_add，可以按照指定的索引（indices），用updates对ref进行更新，不过这里的索引可以是对多维张量的索引（注意使用时必须可以在ref中索引到），updates的维度应该和ref中一个indices对应的张量维度相同，scatter_nd_sub（相减，ref[index]-updates[index]），scatter_nd_update（用updates的值替换，不做加减法）同理

有点饶，看代码比较明显：

import tensorflow as tf
import numpy as np
ref = tf.Variable([[[1, 2, 3, 4],[5, 6, 7, 8],
                    [1, 2, 3, 4],[5, 6, 7, 8]],
                   [[10,11,12,13],[14,15,16,17],
                    [10,11,12,13],[14,15,16,17]]
                  ])
indices = tf.constant([[1,1]])# 表示对应的索引，ref[1,1]=[14,15,16,17]
updates = tf.constant([[-14,-15,-16,-17]])# 要更新的值，相加会全变成0
add = tf.scatter_nd_add(ref, indices, updates)
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    print(sess.run(add))
output:
[[[ 1  2  3  4]
  [ 5  6  7  8]
  [ 1  2  3  4]
  [ 5  6  7  8]]
 [[10 11 12 13]
  [ 0  0  0  0]
  [10 11 12 13]
  [14 15 16 17]]]

tf.Variable()

创建变量，在初始化值中指定数据维度和内容，其他东西有点多....，先关注简易版

必须指定初始化

__init__(
    initial_value=None,
    trainable=None,
    collections=None,
    validate_shape=True,
    caching_device=None,
    name=None,
    variable_def=None,
    dtype=None,
    expected_shape=None,
    import_scope=None,
    constraint=None,
    use_resource=None,
    synchronization=tf.VariableSynchronization.AUTO,
    aggregation=tf.VariableAggregation.NONE,
    shape=None
)

import tensorflow as tf
import numpy as np
a1 = tf.Variable(tf.random_normal(shape=[2,3], mean=0, stddev=1), name='a1')
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    print(sess.run(a1))
output:
[[ 0.1626461  1.4406685  0.2573926]
 [-1.911016  -0.6992628 -0.8095152]]

tf.get_variable()

获取已存在的变量，如果不存在，就新建一个，需要指定名称，可以用于参数共享

tf.get_variable(
    name,
    shape=None,
    dtype=None,
    initializer=None,
    regularizer=None,
    trainable=None,
    collections=None,
    caching_device=None,
    partitioner=None,
    validate_shape=True,
    use_resource=None,
    custom_getter=None,
    constraint=None,
    synchronization=tf.VariableSynchronization.AUTO,
    aggregation=tf.VariableAggregation.NONE
)

tf.variable_scope()

tf.variable_scope()与tf.get_variable()配合，用于实现变量共享

tf.name_scope()

tf.name_scope不会对tf.get_variable()创建的变量产生影响，用于管理命名空间

对于使用tf.Variable来说，tf.name_scope和tf.variable_scope功能一样，都是给变量加前缀，相当于分类管理，模块化。

对于tf.get_variable来说，tf.name_scope对其无效，也就是说tf认为当你使用tf.get_variable时，你只归属于tf.variable_scope来管理共享与否。

name_scope为命名域，variable_scope为变量域，

import tensorflow as tf

# x = tf.get_variable('x',shape=[32,32],initializer=tf.initializers.random_normal)
# y = tf.get_variable('y',shape=[32,32],initializer=tf.initializers.random_normal)

with tf.name_scope('name_scope'):
    x = tf.get_variable('x',shape=[32,32],initializer=tf.initializers.random_normal)
    y = tf.get_variable('y',shape=[32,32],initializer=tf.initializers.random_normal)
    print(x.name,y.name)

with tf.variable_scope('variable_scope'):
    x = tf.get_variable('x',shape=[32,32],initializer=tf.initializers.random_normal)
    y = tf.get_variable('y',shape=[32,32],initializer=tf.initializers.random_normal)
    print(x.name,y.name)

output:
x:0 y:0
variable_scope/x:0 variable_scope/y:0

import tensorflow as tf
x = tf.placeholder(dtype=tf.float32,shape=(None,32,32,3))
with tf.variable_scope('extra/conv1'):  
    conv1 = tf.layers.conv2d(x,64,(3,3))
    conv1 = tf.layers.batch_normalization(conv1)
    print(conv1)
def extra(x):
    with tf.variable_scope('extra',reuse=tf.AUTO_REUSE):
        with tf.variable_scope('conv1'):  
            conv1 = tf.layers.conv2d(x,64,(3,3))
            conv1 = tf.layers.batch_normalization(conv1)
            print(conv1)
        with tf.variable_scope('conv2'):  
            conv2 = tf.layers.conv2d(conv1,64,(3,3))
            conv2 = tf.layers.batch_normalization(conv2)
            print(conv2)
    with tf.variable_scope('fc',reuse=tf.AUTO_REUSE):
        fc = tf.layers.flatten(conv2)
        logits = tf.layers.dense(fc,10)
        print(logits)
    return logits
y1 = extra(x)  
y2 = extra(x)
vars = tf.trainable_variables()
for k,v in enumerate(vars):
    print(k,v)
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
    bn = tf.get_variable(name='extra1/conv2/batch_normalization/gamma',shape=(64,))
    print(bn)

output:
Tensor("extra/conv1/batch_normalization/FusedBatchNorm:0", shape=(?, 30, 30, 64), dtype=float32)
Tensor("extra/conv1_1/batch_normalization/FusedBatchNorm:0", shape=(?, 30, 30, 64), dtype=float32)
Tensor("extra/conv2/batch_normalization/FusedBatchNorm:0", shape=(?, 28, 28, 64), dtype=float32)
Tensor("fc/dense/BiasAdd:0", shape=(?, 10), dtype=float32)
Tensor("extra_1/conv1/batch_normalization/FusedBatchNorm:0", shape=(?, 30, 30, 64), dtype=float32)
Tensor("extra_1/conv2/batch_normalization/FusedBatchNorm:0", shape=(?, 28, 28, 64), dtype=float32)
Tensor("fc_1/dense/BiasAdd:0", shape=(?, 10), dtype=float32)
0 <tf.Variable 'extra/conv1/conv2d/kernel:0' shape=(3, 3, 3, 64) dtype=float32_ref>
1 <tf.Variable 'extra/conv1/conv2d/bias:0' shape=(64,) dtype=float32_ref>
2 <tf.Variable 'extra/conv1/batch_normalization/gamma:0' shape=(64,) dtype=float32_ref>
3 <tf.Variable 'extra/conv1/batch_normalization/beta:0' shape=(64,) dtype=float32_ref>
4 <tf.Variable 'extra/conv2/conv2d/kernel:0' shape=(3, 3, 64, 64) dtype=float32_ref>
5 <tf.Variable 'extra/conv2/conv2d/bias:0' shape=(64,) dtype=float32_ref>
6 <tf.Variable 'extra/conv2/batch_normalization/gamma:0' shape=(64,) dtype=float32_ref>
7 <tf.Variable 'extra/conv2/batch_normalization/beta:0' shape=(64,) dtype=float32_ref>
8 <tf.Variable 'fc/dense/kernel:0' shape=(50176, 10) dtype=float32_ref>
9 <tf.Variable 'fc/dense/bias:0' shape=(10,) dtype=float32_ref>
<tf.Variable 'extra1/conv2/batch_normalization/gamma:0' shape=(64,) dtype=float32_ref>

为啥直接print变量，像extra,fc会添加后缀，成为extra_1,fc_1?

tf.get_collection(key,scope)

从一个集合中取出全部变量，是一个列表

tensorflow允许将变量与一个key关联，然后通过指定key来获取关联的变量，标准的key定义在tf.GraphKeys 中：

常见的key，tf.GraphKeys

GLOBAL_VARIABLES:
LOCAL_VARIABLES
MODEL_VARIABLES
TRAINALBEL_VARIABLES
REGULARIZATION_LOSSES
SUMMARIES
QUEUE_RUNNERS
VARIABLES 
UPDATE_OPS 
TRAIN_OP 
SAVERS 

WEIGHTS
BIASES
ACTIVATIONS

tf.group()

将一个op的list合并成一个op，创建一个新的op

op = tf.group(*[op1,op2])

tf.cond()

tensorflow中的判断语句

tf.cond(
    pred,
    true_fn=None,
    false_fn=None,
    name=None
)
pred:判断条件
true_fn：为真时调用此函数
false_fn：为假时调用此函数

import tensorflow as tf
a=tf.constant(2)    
b=tf.constant(3)    
x=tf.constant(4)    
y=tf.constant(5)    
z = tf.multiply(a, b)# 6
result = tf.cond(x > y, lambda: tf.add(x, z), lambda: tf.subtract(y,z))# 调用减法  
with tf.Session() as session:    
    print(result.eval())
output:-1

tf.clip_by_value()：将t按照指定的最大和最小值进行截断

tf.clip_by_value(
    t,
    clip_value_min,
    clip_value_max,
    name=None
)

weixin_39946534

关注

0
点赞
踩
1

收藏

觉得还不错? 一键收藏
0
评论
tensorflow 按维度相加_TensorFlow一些函数（1）

记录几个tensorflow函数用法tf.unique_with_counts(x,out_idx=tf.dtypes.int32,name=None) x表示一个一维张量，此函数用来统计x中不同元素的信息，可以统计标签信息import tensorflow as tfimport numpy as npx = tf.placeholder(dtype=tf.int32,shape=(10,)...
复制链接

扫一扫