命名空间及变量共享
# coding=utf-8
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt;
with tf.variable_scope('V1') as scope:
a1 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
scope.reuse_variables()
a3 = tf.get_variable('a1')
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print a1.name
print sess.run(a1)
print a3.name
print sess.run(a3)
等同于
with tf.variable_scope('V1'):
a1 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
with tf.variable_scope('V1', reuse=True):
a3 = tf.get_variable('a1')
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print a1.name
print sess.run(a1)
print a3.name
print sess.run(a3)
数学运算
基本数学函数
相同大小Tensor之间的任何算术运算都会将运算应用到元素级。
# 算术操作符:+ - * / %
tf.add(x, y, name=None) # 加法(支持 broadcasting)
tf.subtract(x, y, name=None) # 减法
tf.multiply(x, y, name=None) # 乘法
tf.divide(x, y, name=None) # 浮点除法, 返回浮点数(python3 除法)
tf.mod(x, y, name=None) # 取余
# 幂指对数操作符:^ ^2 ^0.5 e^ ln
tf.pow(x, y, name=None) # 幂次方
tf.square(x, name=None) # 平方
tf.sqrt(x, name=None) # 开根号,必须传入浮点数或复数
tf.exp(x, name=None) # 计算 e 的次方
tf.log(x, name=None) # 以 e 为底,必须传入浮点数或复数
# 取符号、负、倒数、绝对值、近似、两数中较大/小的
tf.negative(x, name=None) # 取负(y = -x).
tf.sign(x, name=None) # 返回 x 的符号
tf.reciprocal(x, name=None) # 取倒数
tf.abs(x, name=None) # 求绝对值
tf.round(x, name=None) # 四舍五入
tf.ceil(x, name=None) # 向上取整
tf.floor(x, name=None) # 向下取整
tf.rint(x, name=None) # 取最接近的整数
tf.maximum(x, y, name=None) # 返回两tensor中的最大值 (x > y ? x : y)
tf.minimum(x, y, name=None) # 返回两tensor中的最小值 (x < y ? x : y)
# 三角函数和反三角函数
tf.cos(x, name=None)
tf.sin(x, name=None)
tf.tan(x, name=None)
tf.acos(x, name=None)
tf.asin(x, name=None)
tf.atan(x, name=None)
# 其它
tf.div(x, y, name=None) # python 2.7 除法, x/y-->int or x/float(y)-->float
tf.truediv(x, y, name=None) # python 3 除法, x/y-->float
tf.floordiv(x, y, name=None) # python 3 除法, x//y-->int
tf.realdiv(x, y, name=None)
tf.truncatediv(x, y, name=None)
tf.floor_div(x, y, name=None)
tf.truncatemod(x, y, name=None)
tf.floormod(x, y, name=None)
tf.cross(x, y, name=None)
tf.add_n(inputs, name=None) # inputs: A list of Tensor objects, each with same shape and type
tf.squared_difference(x, y, name=None)
矩阵函数
# 矩阵乘法(tensors of rank >= 2)
tf.matmul(a, b, transpose_a=False, transpose_b=False, adjoint_a=False, adjoint_b=False, a_is_sparse=False, b_is_sparse=False, name=None)
# 转置,可以通过指定 perm=[1, 0] 来进行轴变换
tf.transpose(a, perm=None, name='transpose')
# 在张量 a 的最后两个维度上进行转置
tf.matrix_transpose(a, name='matrix_transpose')
# Matrix with two batch dimensions, x.shape is [1, 2, 3, 4]
# tf.matrix_transpose(x) is shape [1, 2, 4, 3]
# 求矩阵的迹
tf.trace(x, name=None)
# 计算方阵行列式的值
tf.matrix_determinant(input, name=None)
# 求解可逆方阵的逆,input 必须为浮点型或复数
tf.matrix_inverse(input, adjoint=None, name=None)
# 奇异值分解
tf.svd(tensor, full_matrices=False, compute_uv=True, name=None)
# QR 分解
tf.qr(input, full_matrices=None, name=None)
# 求张量的范数(默认2)
tf.norm(tensor, ord='euclidean', axis=None, keep_dims=False, name=None)
# 构建一个单位矩阵, 或者 batch 个矩阵,batch_shape 以 list 的形式传入
tf.eye(num_rows, num_columns=None, batch_shape=None, dtype=tf.float32, name=None)
# Construct one identity matrix.
tf.eye(2)
==> [[1., 0.],
[0., 1.]]
# Construct a batch of 3 identity matricies, each 2 x 2.
# batch_identity[i, :, :] is a 2 x 2 identity matrix, i = 0, 1, 2.
batch_identity = tf.eye(2, batch_shape=[3])
# Construct one 2 x 3 "identity" matrix
tf.eye(2, num_columns=3)
==> [[ 1., 0., 0.],
[ 0., 1., 0.]]
# 构建一个对角矩阵,rank = 2*rank(diagonal)
tf.diag(diagonal, name=None)
# 'diagonal' is [1, 2, 3, 4]
tf.diag(diagonal) ==> [[1, 0, 0, 0]
[0, 2, 0, 0]
[0, 0, 3, 0]
[0, 0, 0, 4]]
# 其它
tf.diag_part
tf.matrix_diag
tf.matrix_diag_part
tf.matrix_band_part
tf.matrix_set_diag
tf.cholesky
tf.cholesky_solve
tf.matrix_solve
tf.matrix_triangular_solve
tf.matrix_solve_ls
tf.self_adjoint_eig
tf.self_adjoint_eigvals
归约Reduction操作
# 计算输入 tensor 所有元素的和,或者计算指定的轴所有元素的和
tf.reduce_sum(input_tensor, axis=None, keep_dims=False, name=None)
# 'x' is [[1, 1, 1]
# [1, 1, 1]]
tf.reduce_sum(x) ==> 6
tf.reduce_sum(x, 0) ==> [2, 2, 2]
tf.reduce_sum(x, 1) ==> [3, 3]
tf.reduce_sum(x, 1, keep_dims=True) ==> [[3], [3]] # 维度不缩减
tf.reduce_sum(x, [0, 1]) ==> 6
# 计算输入 tensor 所有元素的均值/最大值/最小值/积/逻辑与/或
# 或者计算指定的轴所有元素的均值/最大值/最小值/积/逻辑与/或(just like reduce_sum)
tf.reduce_mean(input_tensor, axis=None, keep_dims=False, name=None)
tf.reduce_max(input_tensor, axis=None, keep_dims=False, name=None)
tf.reduce_min(input_tensor, axis=None, keep_dims=False, name=None)
tf.reduce_prod(input_tensor, axis=None, keep_dims=False, name=None)
tf.reduce_all(input_tensor, axis=None, keep_dims=False, name=None) # 全部满足条件
tf.reduce_any(input_tensor, axis=None, keep_dims=False, name=None) #至少有一个满足条件
-------------------------------------------
# 分界线以上和 Numpy 中相应的用法完全一致
-------------------------------------------
# inputs 为一 list, 计算 list 中所有元素的累计和,
# tf.add(x, y, name=None)只能计算两个元素的和,此函数相当于扩展了其功能
tf.accumulate_n(inputs, shape=None, tensor_dtype=None, name=None)
# Computes log(sum(exp(elements across dimensions of a tensor)))
tf.reduce_logsumexp(input_tensor, axis=None, keep_dims=False, name=None)
# Computes number of nonzero elements across dimensions of a tensor
tf.count_nonzero(input_tensor, axis=None, keep_dims=False, name=None)
Scan
# Compute the cumulative sum of the tensor x along axis
tf.cumsum(x, axis=0, exclusive=False, reverse=False, name=None)
# Eg:
tf.cumsum([a, b, c]) # => [a, a + b, a + b + c]
tf.cumsum([a, b, c], exclusive=True) # => [0, a, a + b]
tf.cumsum([a, b, c], reverse=True) # => [a + b + c, b + c, c]
tf.cumsum([a, b, c], exclusive=True, reverse=True) # => [b + c, c, 0]
# Compute the cumulative product of the tensor x along axis
tf.cumprod(x, axis=0, exclusive=False, reverse=False, name=None)
Segmentation
# Computes the sum/mean/max/min/prod along segments of a tensor
tf.segment_sum(data, segment_ids, name=None)
# Eg:
m = tf.constant([5,1,7,2,3,4,1,3])
s_id = [0,0,0,1,2,2,3,3]
s.run(tf.segment_sum(m, segment_ids=s_id))
>array([13, 2, 7, 4], dtype=int32)
tf.segment_mean(data, segment_ids, name=None)
tf.segment_max(data, segment_ids, name=None)
tf.segment_min(data, segment_ids, name=None)
tf.segment_prod(data, segment_ids, name=None)
# 其它
tf.unsorted_segment_sum
tf.sparse_segment_sum
tf.sparse_segment_mean
tf.sparse_segment_sqrt_n
分割
tf.split(value, num_or_size_splits, axis=0, num=None, name='split')
# 'value' is a tensor with shape [5, 30]
# Split 'value' into 3 tensors with sizes [4, 15, 11] along dimension 1
split0, split1, split2 = tf.split(value, [4, 15, 11], 1)
tf.shape(split0) # [5, 4]
tf.shape(split1) # [5, 15]
tf.shape(split2) # [5, 11]
# Split 'value' into 3 tensors along dimension 1
split0, split1, split2 = tf.split(value, num_or_size_splits=3, axis=1)
tf.shape(split0) # [5, 10]
tf.slice(input_, begin, size, name=None)
t = tf.constant([[[1, 1, 1], [2, 2, 2]],
[[3, 3, 3], [4, 4, 4]],
[[5, 5, 5], [6, 6, 6]]])
tf.slice(t, [1, 0, 0], [1, 1, 3]) # [[[3, 3, 3]]]
tf.slice(t, [1, 0, 0], [1, 2, 3]) # [[[3, 3, 3],
# [4, 4, 4]]]
tf.slice(t, [1, 0, 0], [2, 1, 3]) # [[[3, 3, 3]],
# [[5, 5, 5]]]
序列比较与索引提取
# 比较两个 list 或者 string 的不同,并返回不同的值和索引
tf.setdiff1d(x, y, index_dtype=tf.int32, name=None)
# 返回 x 中的唯一值所组成的tensor 和原 tensor 中元素在现 tensor 中的索引
tf.unique(x, out_idx=None, name=None)
# x if condition else y, condition 为 bool 类型的,可用tf.equal()等来表示
# x 和 y 的形状和数据类型必须一致
tf.where(condition, x=None, y=None, name=None)
# 返回沿着坐标轴方向的最大/最小值的索引
tf.argmax(input, axis=None, name=None, output_type=tf.int64)
tf.argmin(input, axis=None, name=None, output_type=tf.int64)
# x 的值当作 y 的索引,range(len(x)) 索引当作 y 的值
# y[x[i]] = i for i in [0, 1, ..., len(x) - 1]
tf.invert_permutation(x, name=None)
# 其它
tf.edit_distance
常用函数
tf.concat
把一组向量从某一维上拼接起来,很向numpy中的Concatenate,官网例子:
t1 = [[1, 2, 3], [4, 5, 6]]
t2 = [[7, 8, 9], [10, 11, 12]]
tf.concat([t1, t2], 0) ==> [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]]
tf.concat([t1, t2], 1) ==> [[1, 2, 3, 7, 8, 9], [4, 5, 6, 10, 11, 12]]
# tensor t3 with shape [2, 3]
# tensor t4 with shape [2, 3]
tf.shape(tf.concat([t3, t4], 0)) ==> [4, 3]
如果是list类型的话也是可以的,只要是形似Tensor,最后tf.concat返回的还是Tensor类型
tf.gather
类似于数组的索引,可以把向量中某些索引值提取出来。只适合在一维的情况下使用。
import tensorflow as tf
a = tf.Variable([[1,2,3,4,5], [6,7,8,9,10], [11,12,13,14,15]])
index_a = tf.Variable([0,2])
b = tf.Variable([1,2,3,4,5,6,7,8,9,10])
index_b = tf.Variable([2,4,6,8])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(tf.gather(a, index_a)))
print(sess.run(tf.gather(b, index_b)))
# [[ 1 2 3 4 5]
# [11 12 13 14 15]]
# [3 5 7 9]
tf.gather_nd
同上,但允许在多维上进行索引。
tf.greater
判断函数。首先张量x和张量y的尺寸要相同,输出的tf.greater(x, y)也是一个和x,y尺寸相同的张量。如果x的某个元素比y中对应位置的元素大,则tf.greater(x, y)对应位置返回True,否则返回False。与此类似的函数还有tf.less、tf.greater_equal。
tf.cast
a = tf.constant([0, 2, 0, 4, 2, 2], dtype='int32')
print(a)
# <tf.Tensor 'Const_1:0' shape=(6,) dtype=int32>
b = tf.cast(a, 'float32')
print(b)
# <tf.Tensor 'Cast:0' shape=(6,) dtype=float32>
tf.expand_dims & tf.squeeze
增加 / 压缩张量的维度。
a = tf.constant([0, 2, 0, 4, 2, 2], dtype='int32')
print(a)
# <tf.Tensor 'Const_1:0' shape=(6,) dtype=int32>
b = tf.expand_dims(a, 0)
print(b)
# <tf.Tensor 'ExpandDims:0' shape=(1, 6) dtype=int32>
print(tf.squeeze(b, 0))
# <tf.Tensor 'Squeeze:0' shape=(6,) dtype=int32>
Tensor的随机存取和遍历
只要能事先知道tensor的size,都可以通过python的循环来对tensor的entry遍历处理。
import tensorflow as tf
data=tf.constant([[1,2,3],[4,5,6]])
aa=data*1
size=aa.get_shape()
sum=tf.convert_to_tensor(0)
for i in range(2):
for j in range(2):
sum=sum+data[i][j]
with tf.Session() as sess:
print(sess.run([sum, size]))
但在tensor size只有在运行的时候才能确定时,比如输入不同尺寸的图片,不同数量的bounding box,就没把发在定义graph的时候就确定个数,这时只有使用tf.while_loop。