从0开始学pytorch【4】--维度变换、拼接与拆分

女爷i 2023-10-04 23:58 1阅读 0赞

#### 从0开始学pytorch【4】--维度变换、拼接与拆分 ####

*  学习内容：
 *  维度变换：
 *  张量拆分与拼接：
 *  小结

## 学习内容： ##

`维度变换、张量拆分与拼接`

*  view/reshape
 *  squeeze/unsqueeze
 *  transpore/t/permute
 *  expand/repeat
 *  cat
 *  stack
 *  split
 *  chunk

--------------------

## 维度变换： ##

1、view

import torch
    a = torch.rand(4,1,28,28)
    print(a.shape)
    print(a.view(4, 28*28))
    print(a.shape)
    b = a.view(4, 28, -1)
    b.view(4, 28,28, -1).shape
    
    '''
    torch.Size([4, 1, 28, 28])
    tensor([[0.2516, 0.4973, 0.2032,  ..., 0.1892, 0.5932, 0.9167],
            [0.9366, 0.3864, 0.8891,  ..., 0.3008, 0.7179, 0.5442],
            [0.5551, 0.8872, 0.8677,  ..., 0.3314, 0.7326, 0.2640],
            [0.2746, 0.3667, 0.0543,  ..., 0.9942, 0.4714, 0.9417]])
    torch.Size([4, 1, 28, 28])
    torch.Size([4, 28, 28, 1])
    '''

2、unsqueeze

插入维度（以增加中括号实现，不改变数据个数多少）

print(a.shape)
    a.unsqueeze(0).shape, a.unsqueeze(-1).shape, a.unsqueeze(4).shape, a.unsqueeze(-4).shape, a.unsqueeze(-5).shape
    
    '''
    torch.Size([4, 1, 28, 28])
    (torch.Size([1, 4, 1, 28, 28]),
     torch.Size([4, 1, 28, 28, 1]),
     torch.Size([4, 1, 28, 28, 1]),
     torch.Size([4, 1, 1, 28, 28]),
     torch.Size([1, 4, 1, 28, 28]))
    '''

直观, 维度体现在，最外层（所有元素）（0）， 最里层（每个元素）（-1）

a = torch.tensor([1.2, 2.3])
    a.unsqueeze(-1), a.unsqueeze(0)
    
    '''
    (tensor([[1.2000],
             [2.3000]]),
     tensor([[1.2000, 2.3000]]))
    '''

维度增加

bias = torch.rand(1, 32, 1, 1)
    f = torch.rand(4, 32, 14, 14)
    print((f+bias).shape)
    
    bias = torch.rand(32)
    bias = bias.unsqueeze(1).unsqueeze(2).unsqueeze(0)
    (f+bias).shape
    
    '''
    torch.Size([4, 32, 14, 14])
    torch.Size([4, 32, 14, 14])
    '''

squeeze,维度删减，体现在中括号的减少，元素个数不变

b = torch.rand(1, 32, 1, 1)
    print(b.squeeze().shape) # 将所有维度为1的都去掉
    print(b.squeeze(0).shape)
    print(b.squeeze(-1).shape)
    print(b.squeeze(1).shape) # 无变化
    b.squeeze(-4).shape
    
    '''
    torch.Size([32])
    torch.Size([32, 1, 1])
    torch.Size([1, 32, 1])
    torch.Size([1, 32, 1, 1])
    torch.Size([32, 1, 1])
    '''

expand,逻辑上复制，实际上内存不改变

a = torch.rand(4, 32, 14, 14)
    b = torch.randn(1, 32, 1, 1)
    print(a.shape, b.shape)
    
    b.expand(4, 32, 14, 14).shape, b.expand(-1, 32, -1, -1).shape, b.shape
    
    '''
    torch.Size([4, 32, 14, 14]) torch.Size([1, 32, 1, 1])
    (torch.Size([4, 32, 14, 14]),
     torch.Size([1, 32, 1, 1]),
     torch.Size([1, 32, 1, 1]))
    '''

repeat, 复制

print(b.shape, 32*32)
    
    b.repeat(4, 32, 1, 1).shape, b.repeat(4, 1, 1,1).shape, b.repeat(4, 1, 32, 32).shape
    
    '''
    torch.Size([1, 32, 1, 1]) 1024
    (torch.Size([4, 1024, 1, 1]),
     torch.Size([4, 32, 1, 1]),
     torch.Size([4, 32, 32, 32]))
    '''

t转置

a = torch.randn(3, 4)
    a.shape, a.t().shape
    
    '''
    (torch.Size([3, 4]), torch.Size([4, 3]))
    '''

transpose维度交换

a = torch.rand(4, 3, 30, 32)
    a.shape, a.transpose(1, 3).shape
    
    '''
    (torch.Size([4, 3, 30, 32]), torch.Size([4, 32, 30, 3]))
    '''

contiguous() 对维度转换后的张量进行复制，必要步骤，否则会改变原始张量，无法作比较

a = torch.rand(4, 3, 30, 32)
    a1 = a.transpose(1, 3).contiguous().view(4, 3*30*32).view(4, 3, 30, 32)
    a2 = a.transpose(1, 3).contiguous().view(4, 3*30*32).view(4, 32, 30, 3).transpose(1,3)
    print(a1.shape, a2.shape)
    
    print(torch.all(torch.eq(a, a1)), torch.all(torch.eq(a, a1)).int())
    print(torch.all(torch.eq(a, a2)), torch.all(torch.eq(a, a2)).int())
    
    '''
    torch.Size([4, 3, 30, 32]) torch.Size([4, 3, 30, 32])
    tensor(False) tensor(0, dtype=torch.int32)
    tensor(True) tensor(1, dtype=torch.int32)
    '''

permute, 按照序号，直接调换各维度位置

a = torch.randn(4, 3, 28, 30)
    a.shape, a.permute(3, 0, 2, 1).shape
    
    '''
    (torch.Size([4, 3, 28, 30]), torch.Size([30, 4, 28, 3]))
    '''

broadcasting ， 同tensorflow, 都是broadcast\_to, 从最里层的维度开始对齐扩张， 从最后一个维度开始。

x = torch.tensor([1, 2, 3])
    print(x, x.shape, torch.broadcast_to(x, (6, 3)))
    
    '''
    tensor([1, 2, 3]) torch.Size([3]) tensor([[1, 2, 3],
            [1, 2, 3],
            [1, 2, 3],
            [1, 2, 3],
            [1, 2, 3],
            [1, 2, 3]])
    '''

torch.broadcast\_tensors 返回两个值，变换后的两个值

x = torch.rand(4, 2)
    bias = torch.randn(1, 2)
    
    # torch.broadcast_tensors 返回两个值，变换后的两个值
    bias1 = torch.broadcast_tensors(bias, x)
    x,  bias,  bias1
    
    '''
    (tensor([[0.5489, 0.0084],
             [0.2405, 0.0769],
             [0.2159, 0.5728],
             [0.2529, 0.7946]]),
     tensor([[0.3781, 0.9359]]),
     (tensor([[0.3781, 0.9359],
              [0.3781, 0.9359],
              [0.3781, 0.9359],
              [0.3781, 0.9359]]),
      tensor([[0.5489, 0.0084],
              [0.2405, 0.0769],
              [0.2159, 0.5728],
              [0.2529, 0.7946]])))
    '''

--------------------

## 张量拆分与拼接： ##

cat拼接

a = torch.rand(4, 32, 8)
    b = torch.rand(5, 32, 8)
    print(torch.cat([a, b], dim=0).shape)
    
    a1 = torch.rand(4, 3, 32, 32)
    a2 = torch.rand(5, 3, 32, 32)
    print(torch.cat([a1, a2], dim=0).shape)
    
    a1 = torch.rand(4, 1, 32, 32)
    a2 = torch.rand(4, 6, 32, 32)
    print(torch.cat([a1, a2], dim=1).shape)
    
    a1 = torch.rand(4, 3, 32, 32)
    a2 = torch.rand(4, 3, 32, 3)
    print(torch.cat([a1, a2], dim=-1).shape)
    
    '''
    torch.Size([9, 32, 8])
    torch.Size([9, 3, 32, 32])
    torch.Size([4, 7, 32, 32])
    torch.Size([4, 3, 32, 35])
    '''

stack, 新增维度，开关. 要求两个张量维度一致

a1 = torch.rand(4, 3, 16, 32)
    a2 = torch.randn(4, 3, 16, 32)
    print(torch.cat([a1, a2], dim=2).shape)
    
    print(torch.stack([a1, a2], dim=2).shape)
    
    a = torch.rand(32, 8)
    b = torch.rand(32, 8)
    
    torch.stack([a, b], dim=0).shape
    
    '''
    torch.Size([4, 3, 32, 32])
    torch.Size([4, 3, 2, 16, 32])
    torch.Size([2, 32, 8])
    '''

split 按照维度和元素个数 切割

a = torch.rand(32, 8)
    b = torch.rand(32, 8)
    
    c = torch.stack([a, b], dim=0)
    print(c.shape)
    
    aa, bb = c.split([1, 1], dim=0)
    print(aa.shape, bb.shape)
    
    aa, bb = c.split([7, 1], dim=-1)
    print(aa.shape, bb.shape)
    
    aa, bb = c.split(1, dim=0)
    print(aa.shape, bb.shape)
    
    c = torch.cat([a, b], dim=0)
    print(c.shape)
    
    aa, bb = c.split([1,63], dim=0)
    print(aa.shape, bb.shape)
    
    '''
    torch.Size([2, 32, 8])
    torch.Size([1, 32, 8]) torch.Size([1, 32, 8])
    torch.Size([2, 32, 7]) torch.Size([2, 32, 1])
    torch.Size([1, 32, 8]) torch.Size([1, 32, 8])
    torch.Size([64, 8])
    torch.Size([1, 8]) torch.Size([63, 8])
    '''

chunk 以及填充、对数

c = torch.randn(2, 32, 8)
    print(c.shape)
    
    aa, bb = c.chunk(2, dim=0)
    print(aa.shape, bb.shape)
    
    a = torch.full([4], 10)
    print(torch.log10(a),torch.log(a),torch.log2(a))
    
    '''
    torch.Size([2, 32, 8])
    torch.Size([1, 32, 8]) torch.Size([1, 32, 8])
    tensor([1., 1., 1., 1.]) tensor([2.3026, 2.3026, 2.3026, 2.3026]) tensor([3.3219, 3.3219, 3.3219, 3.3219])
    '''

--------------------

## 小结 ##

*  view/reshape
 *  squeeze/unsqueeze
 *  transpore/t/permute
 *  expand/repeat
 *  cat
 *  stack
 *  split
 *  chunk

--------------------