矩陣操作

#對於2-D
#所有的reduce_...，如果不加axis的話，都是對整個矩陣進行運算
tf.reduce_sum(a, 1） #對axis1
tf.reduce_mean(a,0) #每列均值

第二個引數是axis，如果為0的話，res[i]=∑ja[j,i]$res[i]=\sum _{j}a[j,i]$即（res[i]=∑a[:,i]$res[i]=\sum a[:,i]$）， 如果是1的話，res[i]=∑ja[i,j]$res[i]=\sum _{j}a[i,j]$
NOTE:返回的都是行向量,（axis等於幾，就是對那維操作,i.e.:沿著那維操作, 其它維度保留）

#關於concat，可以用來進行降維 3D->2D , 2D->1D
 

tf.concat(concat_dim, data)
#arr = np.zeros([2,3,4,5,6])
In [6]: arr2.shape
Out[6]: (2, 3, 4, 5)
In [7]: np.concatenate(arr2, 0).shape
Out[7]: (6, 4, 5)   :(2*3, 4, 5)
In [9]: np.concatenate(arr2, 1).shape
Out[9]: (3, 8, 5)   :(3, 2*4, 5)
#tf.concat()
t1 = [[1, 2, 3], [4, 5, 6]]
t2 = [[7, 8, 9], [10, 11
 
, 12]]
# 將t1, t2進行concat，axis為0，等價於將shape=[2, 2, 3]的Tensor concat成
#shape=[4, 3]的tensor。在新生成的Tensor中tensor[:2,:]代表之前的t1
#tensor[2:,:]是之前的t2
tf.concat(0, [t1, t2]) ==> [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]]

# 將t1, t2進行concat，axis為1，等價於將shape=[2, 2, 3]的Tensor concat成
#shape=[2, 6]的tensor。在新生成的Tensor中tensor[:,:3]代表之前的t1
 

#tensor[:,3:]是之前的t2
tf.concat(1, [t1, t2]) ==> [[1, 2, 3, 7, 8, 9], [4, 5, 6, 10, 11, 12]]

concat是將list中的向量給連線起來，axis表示將那維的資料連線起來，而其他維的結構保持不變

#squeeze 降維 維度為1的降掉
tf.squeeze(arr, [])
降維， 將維度為1 的降掉
arr = tf.Variable(tf.truncated_normal([3,4,1,6,1], stddev=0.1))
arr2 = tf.squeeze(arr, [2,4])
arr3 = tf.squeeze(arr) #降掉所以是1的維

#split
tf.split(split_dim, num_split, value, name='split')
# 'value' is a tensor with shape [5, 30]
# Split 'value' into 3 tensors along dimension 1
split0, split1, split2 = tf.split(1, 3, value)
tf.shape(split0) ==> [5, 10]

#embedding
mat = np.array([1,2,3,4,5,6,7,8,9]).reshape((3,-1))
ids = [[1,2], [0,1]]
res = tf.nn.embedding_lookup(mat, ids)
res.eval()
array([[[4, 5, 6],
        [7, 8, 9]],

       [[1, 2, 3],
        [4, 5, 6]]])

#擴充套件維度，如果想用廣播特性的話，經常會用到這個函式
# 't' is a tensor of shape [2]
#一次擴充套件一維
shape(tf.expand_dims(t, 0)) ==> [1, 2]
shape(tf.expand_dims(t, 1)) ==> [2, 1]
shape(tf.expand_dims(t, -1)) ==> [2, 1]
# 't2' is a tensor of shape [2, 3, 5]
shape(tf.expand_dims(t2, 0)) ==> [1, 2, 3, 5]
shape(tf.expand_dims(t2, 2)) ==> [2, 3, 1, 5]
shape(tf.expand_dims(t2, 3)) ==> [2, 3, 5, 1]

tf.slice()

tf.slice(input_, begin, size, name=None)
先看例子

import tensorflow as tf
import numpy as np
sess = tf.InteractiveSession()
a = np.array([[1,2,3,4,5],[4,5,6,7,8],[9,10,11,12,13]])
tf.slice(a,[1,2],[-1,2]).eval()

#array([[ 6,  7],
#       [11, 12]])

理解tf.slice()最好是從返回值上去理解，現在假設input的shape是[a1, a2, a3], begin的值是[b1, b2, b3],size的值是[s1, s2, s3],那麼tf.slice()返回的值就是 input[b1:b1+s1, b2:b2+s2, b3:b3+s3]。
如果 si=−1$s_i=-1$ ，那麼 返回值就是 input[b1:b1+s1,..., bi: ,...]

注意：input[1:2] 取不到input[2]

tf.stack()

tf.stack(values, axis=0, name=’stack’)

將 a list of R 維的Tensor堆成 R+1維的Tensor。
Given a list of length N of tensors of shape (A, B, C);
if axis == 0 then the output tensor will have the shape (N, A, B, C)

這時 res[i,:,:,:] 就是原 list中的第 i 個 tensor

. if axis == 1 then the output tensor will have the shape (A, N, B, C).

這時 res[:,i,:,:] 就是原list中的第 i 個 tensor

Etc.

# 'x' is [1, 4]
# 'y' is [2, 5]
# 'z' is [3, 6]
stack([x, y, z]) => [[1, 4], [2, 5], [3, 6]]  # Pack along first dim.
stack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]]

tf.gather()

tf.gather(params, indices, validate_indices=None, name=None)

indices must be an integer tensor of any dimension (usually 0-D or 1-D). Produces an output tensor with shape indices.shape + params.shape[1:]

# Scalar indices, 會降維
output[:, ..., :] = params[indices, :, ... :]

# Vector indices
output[i, :, ..., :] = params[indices[i], :, ... :]

# Higher rank indices，會升維
output[i, ..., j, :, ... :] = params[indices[i, ..., j], :, ..., :]

tf.pad

tf.pad(tensor, paddings, mode="CONSTANT", name=None)

• tensor: 任意shape的tensor，維度 Dn
• paddings: [Dn, 2] 的 Tensor, Padding後tensor的某維上的長度變為padding[D,0]+tensor.dim_size(D)+padding[D,1]
• mode: CONSTANT表示填0, REFLECT表示反射填充,SYMMETRIC表示對稱填充。

tf.scatter_nd

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

# 文件中廢話那麼多，可以總結成兩個式子
# res[*indice[i,j,..,z], ...] = updates[i,j,..,z,...]
# len([i,j,..,z]) = indice.rank-2
# 在這裡 rank 表示 幾維， a=[1,2,3], a.rank=1,  b = [[1,2], [2,3], [3,4]], b.rank=2