pytorch 学习笔记——维度变换

view/reshape

In [13]: a=torch.rand(4,1,28,28)

In [14]: a.shape
Out[14]: torch.Size([4, 1, 28, 28])

In [15]: a.view(4, 28*28)
Out[15]: tensor([[0.9020, 0.5015, 0.1010, ..., 0.4946, 0.1665, 0.5352],
				[0.3386, 0.5002, 0.7858, ..., 0.1595, 0.9348, 0.3489],
				[0.6014, 0.9255, 0.4829, ..., 0.4963, 0.1805, 0.7925],
				[0.8898, 0.2387, 0.7560, ..., 0.1687, 0.3041, 0.3767]])
//view()前后操作需要size相同
//此处合并了后三个通道

In [16]: a.view(4,28*28).shape
Out[16]: torch.Size([4, 784])

In [17]: a.view(4*28, 28).shape
0ut[17]: torch.Size([112, 28])
//此处合并了前三个通道

In [18]: a.view(4*1, 28, 28).shape 
Out[18]: torch.Size([4, 28, 28])
//此处合并了前两个通道

In [19]: b=a.view(4, 784)
In [20]: b.view(4, 28, 28, 1) # Logic Bug
//b就会丢失掉原来a存储的维度信息

squeeze/unsqueeze

unsqueeze

In [34]: a.shape
Out[34]: torch.Size([4, 1, 28, 28])

In [35]: a.unsqueeze(0).shape
Out[35]: torch.Size([1, 4, 1, 28, 28])
//在0索引前插入，即第一个

In [36]: a.unsqueeze(-1).shape
Out[36]: torch.Size([4, 1, 28, 28, 1])
//在-1索引后插入，即最后一个
//正数就是在之前插入，负数就是在之后插入

In [37]: a.unsqueeze(4).shape
Out[37]: torch.Size([4, 1, 28, 28, 1])

In [38]: a.unsqueeze(-4).shape
Out[38]: torch.Size([4, 1, 1, 28, 28])

In [39]: a. unsqueeze(-5).shape
Out[39]: torch.Size([1, 4, 1, 28, 28])

In [40] : a.unsqueeze(5).shape
RuntimeError: Dimension out of range
(expected to be in range of [-5, 4], but got 5)

In [46]: a=torch.tensor([1.2,2.3])

In [47]: a.unsqueeze(-1)
Out[47]: tensor([[1.2000], [2.3000]])

In [49]: a.unsqueeze(0)
Out[49]: tensor([[1.2000, 2.3000]])

squeeze

In [60]: b.shape
0ut[60]: torch.size([1, 32, 1, 1])

In [61]: b.squeeze().shape
0ut[61]: torch.Size([32])

In [62]: b.squeeze(0).shape
Out[62]: torch.Size([32, 1, 1])

In [63]: b.squeeze(-1).shape
0ut[63]: torch.Size([1, 32, 1])

In [64]: b.squeeze(1).shape
Out[64]: torch.Size([1, 32, 1, 1])

In [65]: b.squeeze(-4).shape
0ut[65]: torch.Size([32, 1, 1])

expand/repeat

expand
只扩展，不添加数据
使用前后需要维度数相同，原需扩展维度的张量必须为1

In [68]: a=torch.rand(4, 32, 14, 14)

In [73]: b.shape
Out[73]: torch.size([1, 32, 1, 1])

In [70] : b.expand(4,32,14,14).shape
0ut[70] : torch.Size([4, 32, 14, 14])
//expand()参数为目标扩展维度张量

In [72]: b.expand(-1,32,-1,-1).shape
Out[72]: torch.Size([1, 32, 1, 1])
//-1为希望该维度的张量保持原状

In [71]:b.expand(-1,32,-1,-4).shape
Out[71]: torch.Size([1, 32, 1, -4])
//-4为输入错误，也会修改为-4但是没有实质含义

repeat
扩展，而且有数据，但与之前的相同

In [74]: b.shape
Out[74]: torch.Size([1, 32, 1, 1])

In [75]: b.repeat(4,32,1,1).shape 
Out[75]: torch.Size([4, 1024, 1, 1] )
//repeat()内参数为拷贝的次数

In [76]: b.repeat(4,1,1,1).shape
Out[76]: torch.Size([4, 32, 1, 1])

In [77]: b.repeat(4,1,32,32).shape
Out[77]: torch.Size([4, 32, 32, 32])

.t()

将矩阵变为转置矩阵

transpose()

将目标的两个维度进行交换