Torch learning notes (2) -- 11 common operation modes of tensor

import torch
import numpy as np

tensor = torch.ones((3, 3))
print(tensor)

# tensor Splicing 
# cat() take tensor By dimension dim Splicing .tensors= Tensor sequence .dim= Dimensions to be spliced 
t_cat_0 = torch.cat([tensor, tensor], dim=0)  # tensor In the 0 Dimension self splicing , The first 0 The dimension is equivalent to （x,y） Of x Splices , So the output is （6,3）
print(t_cat_0, t_cat_0.shape)
t_cat_1 = torch.cat([tensor, tensor], dim=1)  # tensor In the 1 Dimension self splicing , The first 1 The dimension is equivalent to （x,y） Of y Splices , So the output is （3,6）
print(t_cat_1, t_cat_1.shape)
# stack() In the created dimension dim Splice on .tensors= Tensor sequence .dim= Dimensions to be spliced 
t_stack_2 = torch.stack([tensor, tensor, tensor], dim=2)  # tensor In the created page 2 Dimension self splicing , If the selected dimension is an existing dimension , Then the existing dimension is directed to +1 Direction backward 
print(t_stack_2, t_stack_2.shape)

# tensor segmentation 
# chunk() take tensor In the specified dim Upper average segmentation  input = tensor ; chunks =  Number of cuts  ; dim =  Dimension to be cut ,return tensor_list
# war： If it's not divisible , The last one tensor Smaller than others tensor
t_chunk_1_2 = torch.chunk(tensor, dim=1, chunks=2)  #  For two-dimensional 3*3tensor The first dimension is segmented , cut 2
for i, t in enumerate(t_chunk_1_2):
    print(i, t, t.shape)
# split() take tensor By dimension dim segmentation ,tensor Tensor to be cut ,split_szie_or_sections:int Indicates the length of each piece ,list According to the said list segmentation .dim =  Dimension to be cut ,return tensor_list
t_split_1_2 = torch.split(tensor, 2, dim=1)  #  For two-dimensional 3*3tensor The first dimension is segmented , The length is 2, here split_szie_or_sections Is an integer 
for i, t in enumerate(t_split_1_2):
    print(i, t, t.shape)
t_split_1_2 = torch.split(tensor, [1, 2],
                          dim=1)  #  For two-dimensional 3*3tensor The first dimension is segmented , The length is 2, here split_szie_or_sections As the list ,list_sum = tensor_dim
for i, t in enumerate(t_split_1_2):
    print(i, t, t.shape)

# tensor Indexes 
# index_select() In dimension dim On , Press index Index data ,return：index For drinking data splicing tensor,input= To index tensor,dim= To index dimensions ,index= To index serial numbers 
tensor = torch.randint(0, 6, size=(3, 3))
t_select = torch.select(tensor, dim=0, index=1)
print(tensor, '\n', t_select)
# masked_select() Press mask Medium True Index ,return： One dimensional qualified position tensor,input= To index tensor,mask= And input Of the same shape bool_tensor # It is often used to filter data 
mask = tensor.ge(5)  # ge Indicates greater than or equal bool,le Less than or equal to ,lt Say less than 
t_maks = torch.masked_select(tensor, mask)
print(t_maks)

# tensor Transformation 
# reshape() Transformation tensor shape  input Indicates input tensor shape Express tensor shape 
#  Be careful ： When tensor When continuous in memory , new tensor And input share id
t_reshape = torch.reshape(tensor, shape=(9, 1))
print(t_reshape)
# transpose() In exchange for tensor Two dimensions of  input： To exchange tensor dim0,dim1
t_transpose = torch.transpose(tensor, dim0=0, dim1=1)
print(t_transpose)
# t() A two-dimensional tensor Transposition , Equivalent to transpose(input,0,1)
t_t = torch.t(tensor)
print(t_t)
# squeeze() The compression length is 1 Dimensions  dim =None It means to remove all items with a length of 1 The shaft , If specified dim, Then if and only if the shaft length is 1 when , Can be removed 
tensor = torch.rand((1, 2, 3, 1))
print(tensor.shape)
t_squeeze = torch.squeeze(tensor)
print(t_squeeze.shape)
# unsqueeze() basis dim Expand dimensions ,dim =  Extended dimensions 
t_unsqueeze = torch.unsqueeze(tensor,dim=1)
print(tensor.shape)
print(t_unsqueeze.shape)