- 1. 张量的创建
- 2. Tensor的索引和切片
- 2.1 直接索引
- 2.2 连续索引
- 2.3 索引+步长
- 2.4 任意的维度
- 2.5 获取指定维度上的指定索引
- 2.6 使用掩码的索引
- 2.7 使用展平的的索引
- 3. 维度变换
- 3.1 shape转换 view/reshape
- 3.2 增加维度 unsqueeze
- 3.3 维度挤压 squeeze
- 3.4 维度扩展 Expand/repeat
- 3.5 tensor转置
- 4. 合并与分割
- 4.1 cat操作
- 4.2 stack操作
- 4.3 split操作
- 4.4 chunk操作
- 5. 统计属性
- 5.1 范数 norm
- 5.2 mean,sum,min,max,prod
- 5.3 top-k k-th
- 6. Tensor的高阶操作
- 6.1 where
- 6.2 gather
import torch import numpy as np # Import from numpy a1 = np.array([2, 3.3]) b1 = torch.from_numpy(a1) a2 = np.ones([2,3]) b2 = torch.from_numpy(a) # Import from List a3 = torch.tensor([2., 3.2]) a4 = torch.FloatTensor([2., 3.2]) # set default type a5 = torch.tensor([1.2, 3]).type() torch.set_default_tensor_type(torch.DoubleTensor) a6 = tensor.tensor(1.2, 3).type() # rand() 随机0-1初始化 torch.rand(2) # 向量 # tensor([0.2300, 0.7452]) torch.rand(2,3) # 矩阵 # tensor([[0.0214, 0.4409, 0.3778], # [0.8571, 0.4957, 0.7218]]) torch.rand(2,2,3) # 三维 # tensor([[[0.7301, 0.0870, 0.0573], # [0.4576, 0.5049, 0.9410]], # [[0.7045, 0.6702, 0.3119], # [0.1384, 0.9376, 0.9691]]]) # rand_like() 给定tensor的随机0-1初始化 # 创建一个向量 data = torch.tensor([1.1, 2.2, 3.3]) # 随机0-1初始化一个shape和data相同的tensor torch.rand_like(data) # tensor([0.7710, 0.5399, 0.4941]) # randint() 在某一范围内随机整数采样 # 需要指定最小值, 最大值, shape，随机采样包含最小值，不包含最大值 torch.randint(1, 10, size=[2,3]) # tensor([[4, 3, 2], # [5, 2, 2]]) # randn() 随机正态分布 torch.randn(3) # 向量 # tensor([-0.8556, -1.3012, -0.1202]) torch.randn(size=[2,3]) # 矩阵 # tensor([[-0.5486, 1.0590, 2.7441], # [ 0.4896, 0.0592, -0.5964]]) # 全为1的张量 #（1）ones() # 全为1的tensor data = torch.ones(size=[2,3]) # tensor([[1., 1., 1.], # [1., 1., 1.]]) # 和data的shape相同的全为1的tensor torch.ones_like(data) # tensor([[1., 1., 1.], # [1., 1., 1.]]) # 全为0的张量 #（2）zeros() # 全为0的tensor data = torch.zeros(size=[2,3]) # tensor([[0., 0., 0.], # [0., 0., 0.]]) # 和data的shape相同的全为1的tensor torch.zeros_like(data) # tensor([[0., 0., 0.], # [0., 0., 0.]]) # 全为某个值的张量 # 2行3列全为1的tensor torch.full(size=[2,3], fill_value=7.0) # tensor([[7., 7., 7.], # [7., 7., 7.]]) # 初始化一个标量 torch.full(size=[], fill_value=7) # tensor(7) # 初始化一个向量 torch.full(size=[1], fill_value=7) # tensor([7]) # 单位矩阵 # 三行三列，可略写(3,3) torch.eye(3) # tensor([[1., 0., 0.], # [0., 1., 0.], # [0., 0., 1.]]) torch.eye(3,4) # tensor([[1., 0., 0., 0.], # [0., 1., 0., 0.], # [0., 0., 1., 0.]]) # arange() 递增张量 # 从0到10每次递增1，包含0不包含10 torch.arange(0,10) # tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]) # 从0到10每次递增2，包含0不包含10 torch.arange(0, 10, 2) # tensor([0, 2, 4, 6, 8]) # linspace 等分张量 # 从0到10，等分成4份的等差数列，包含0也包含10 torch.linspace(0, 10, 4) # tensor([ 0.0000, 3.3333, 6.6667, 10.0000]) # 0到10等分7份 torch.linspace(0, 10, 7) # tensor([ 0.0000, 1.6667, 3.3333, 5.0000, 6.6667, 8.3333, 10.0000]) # logspace 幂指数递增 # base代表对数的底数，0-10分成11步，计算10^(n) torch.logspace(0, 10, steps=10, base=10) # tensor([1.0000e+00, 1.0000e+01, 1.0000e+02, 1.0000e+03, 1.0000e+04, 1.0000e+05, 1.0000e+06, 1.0000e+07, 1.0000e+08, 1.0000e+09, 1.0000e+10]) # 计算2^(n)，从0到-1分成11步 torch.logspace(0, -1, steps=10, base=2) # tensor([1.0000, 0.9259, 0.8572, 0.7937, 0.7349, 0.6804, 0.6300, 0.5833, 0.5400, 0.5000]) # randperm() 随机打乱 # 生成0-10的随机索引[0,9]之间 torch.randperm(10) # tensor([7, 8, 9, 6, 3, 1, 0, 4, 2, 5]) # 索引打乱后还能配对 index = torch.randperm(10) # 创建两个张量 data1 = torch.arange(0,10) data2 = torch.arange(0,10) * 10 # 索引配对打乱，不改变data1和data2各元素之间的对应关系 data1[index]*10 == data2[index] # tensor([True, True, True, True, True, True, True, True, True, True])
2. Tensor的索引和切片 2.1 直接索引 2.2 连续索引 2.3 索引+步长 2.4 任意的维度 2.5 获取指定维度上的指定索引 2.6 使用掩码的索引 2.7 使用展平的的索引
3. 维度变换 3.1 shape转换 view/reshape
import torch a = torch.randn(4, 1, 28, 28) #假设为MINIST数据集 print(a.shape) b = a.view(4 ,28,28) print(b.shape) c = a.view(4,28*28) #将图像展平 print(c.shape) d = a.view(4*28, 28) #关注所有图像所行 print(d.shape) ''' torch.Size([4, 1, 28, 28]) torch.Size([4, 28, 28]) torch.Size([4, 784]) torch.Size([112, 28]) '''3.2 增加维度 unsqueeze
import torch
a = torch.randn(4, 1, 28, 28) #假设为MINIST数据集
print('a',a.shape)
b = torch.unsqueeze(a, 0) #在0维度前面插入一个维度 没有增加数据
print('b', b.shape)
c = a.unsqueeze(4) #在4维度前面插入一个维度
print('c', c.shape)
d = a.unsqueeze(2) #在第2个维度前面插入一个维度
print('d', d.shape)
e = a.unsqueeze(-1) #在最后-1维度后面插入一个维度
print('e',e.shape)
f = a.unsqueeze(-3) #在-3维度后面之后插入一个维度
print('f',f.shape)
'''
a torch.Size([4, 1, 28, 28])
b torch.Size([1, 4, 1, 28, 28])
c torch.Size([4, 1, 28, 28, 1])
d torch.Size([4, 1, 1, 28, 28])
e torch.Size([4, 1, 28, 28, 1])
f torch.Size([4, 1, 1, 28, 28])
'''
3.3 维度挤压 squeeze
import torch
bias = torch.randn(32)
f = torch.rand(4, 32, 14,14) #bias相当于每个通道上的偏置
# 将bias叠加在f上
b = bias.unsqueeze(0).unsqueeze(2).unsqueeze(3) #[1,32]->[1,32,1]->[1,32,1,1]
print('b', b.shape)
c = b.squeeze() #挤压掉所有为1的维度
print('c', c.shape)
d = b.squeeze(0) #挤压掉0维数据
print('d', d.shape)
e = b.squeeze(1) #挤压掉第1维数据 如果第1维大小不为1,则挤压失败，返回原来的维度
print('e', e.shape)
f = b.squeeze(-4) #挤压第-4维维度
print('f', f.shape)
'''
b torch.Size([1, 32, 1, 1])
c torch.Size([32])
d torch.Size([32, 1, 1])
e torch.Size([1, 32, 1, 1])
f torch.Size([32, 1, 1])
'''
3.4 维度扩展 Expand/repeat
- expend
import torch
bias = torch.randn(32)
f = torch.rand(4, 32, 14,14) #bias相当于每个通道上的偏置
# 将bias叠加在f上
b = bias.unsqueeze(0).unsqueeze(2).unsqueeze(3) #[1,32]->[1,32,1]->[1,32,1,1]
print('b', b.shape)
# c = b.expand(4,10,1,1) #扩展第0个维度 第1个维度不为1，所以该扩展失败（报错），其余维度不变
d = b.expand(4,32,14,14) #扩展所有维度
print(d.shape)
e = b.expand(100,32,1,1)
print(e.shape)
'''
b torch.Size([1, 32, 1, 1])
torch.Size([4, 32, 14, 14])
torch.Size([100, 32, 1, 1])
'''
- repeat
import torch
bias = torch.randn(32)
f = torch.rand(4, 32, 14,14) #bias相当于每个通道上的偏置
# 将bias叠加在f上
b = bias.unsqueeze(0).unsqueeze(2).unsqueeze(3) #[1,32]->[1,32,1]->[1,32,1,1]
print('b', b.shape)
c = b.repeat(4,1,14,14) #[1,32,1,1] -> [14,32,14,14] 拷贝
print('after repeat all dim:', b.shape)
d = b.repeat(10,32,1,1) #[1,32,1,1]->[1*10,32*32,1,1]拷贝第一个维度10次，第二个维度拷贝32次， 其余的拷贝一次
print('repeat dim 0,1:', d.shape)
'''
b torch.Size([1, 32, 1, 1])
after repeat all dim: torch.Size([1, 32, 1, 1])
repeat dim 0,1: torch.Size([10, 1024, 1, 1])
'''
3.5 tensor转置
- .t
import torch
b = torch.randn(3,4)
print('b', b.shape)
c = b.t()
print('b.t()', c.shape)
'''
b torch.Size([3, 4])
b.t() torch.Size([4, 3])
'''
- transpose()
import torch a = torch.randn(4,3,32,32) b = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,3,32,32) #数据污染了 c = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,32,32,3).transpose(1,3) d = a.transpose(1,3).contiguous().view(-1) #使用a.transpose(1,3).view(-1)会报错， 必须要加一个contiguous()函数
- permute转置
#将tensor图像数据转换为numpy图像数据 import torch a = torch.randn(4,3,32,32) print(a.shape) b = a.transpose(1,3).transpose(1,2) #[b,c,h,w]->[b,w,h,c]->[b,h,w,c] print(b.shape) c = a.permute(0,2,3,1) print(c.shape) ''' torch.Size([4, 3, 32, 32]) torch.Size([4, 32, 32, 3]) torch.Size([4, 32, 32, 3]) '''
4. 合并与分割 4.1 cat操作
#tensor拼接 import torch a = torch.rand(4,32,8) b = torch.rand(5,32,8) c = torch.cat((a,b),dim=0) print(c.shape) ''' torch.Size([9, 32, 8]) '''4.2 stack操作
#tensor堆叠 import torch a = torch.rand(4,32,8) b = torch.rand(4,32,8) c = torch.stack((a,b),dim=0) print(c.shape) ''' torch.Size([2, 4, 32, 8]) '''4.3 split操作
#tensor拆分 import torch a = torch.rand(4,3,3) aa,bb = torch.split(a,[1,3], dim=0) print(aa.shape) print(bb.shape) cc,dd = torch.split(a, 2, dim = 0) #在0维度上按照单元长度为2进行拆分 print(cc.shape) print(dd.shape) ''' torch.Size([1, 3, 3]) torch.Size([3, 3, 3]) torch.Size([2, 3, 3]) torch.Size([2, 3, 3]) '''4.4 chunk操作
#tensor拆分 import torch a = torch.rand(4,3,3) bb,cc,dd,ee = torch.chunk(a,4,dim=0) #将零维度拆分为4个块 print(bb.shape) print(cc.shape) print(dd.shape) print(ee.shape) ''' torch.Size([1, 3, 3]) torch.Size([1, 3, 3]) torch.Size([1, 3, 3]) torch.Size([1, 3, 3]) '''
5. 统计属性 5.1 范数 norm
import torch
a = torch.full([8],1, dtype = torch.float)
b = a.view(2,4)
c = a.view(2,2,2)
print(a.norm(1), b.norm(1), c.norm(1))
print(a.norm(2), b.norm(2), c.norm(2))
print(b.norm(1, dim = 1))
print(c.norm(2, dim = 0))
'''
tensor(8.) tensor(8.) tensor(8.)
tensor(2.8284) tensor(2.8284) tensor(2.8284)
tensor([4., 4.])
tensor([[1.4142, 1.4142],
[1.4142, 1.4142]])
'''
5.2 mean,sum,min,max,prod
import torch a = torch.arange(8).view(2,4).float() # mean,max,min等如果没有指定维度，则会将tensor展平，然后再统计 print(a.min(), a.max()) print(a.sum(), a.mean()) print(a.prod()) print(a.argmax(), a.argmin()) ''' tensor(0.) tensor(7.) tensor(28.) tensor(3.5000) tensor(0.) tensor(7) tensor(0) '''
import torch
t = torch.arange(8).reshape(2,4).float()
print('**Caculate after flatten..')
print(t.max())
print(t.min())
print(t.sum())
print(t.mean())
print(t.prod())
print('**Get the position of the max/min elements on all dim..')
print(t.argmax())
print(t.argmin())
# dim指定维度
print('**Gaculate on special dim..')
print(t.mean(1))
print(t.max(1))
print(t.min(0))
print(t.sum(0))
print(t.prod(1))
'''
**Caculate after flatten..
tensor(7.)
tensor(0.)
tensor(28.)
tensor(3.5000)
tensor(0.)
**Get the position of the max/min elements on all dim..
tensor(7)
tensor(0)
**Gaculate on special dim..
tensor([1.5000, 5.5000])
torch.return_types.max(
values=tensor([3., 7.]),
indices=tensor([3, 3]))
torch.return_types.min(
values=tensor([0., 1., 2., 3.]),
indices=tensor([0, 0, 0, 0]))
tensor([ 4., 6., 8., 10.])
tensor([ 0., 840.])
'''
import torch
t = torch.arange(8).reshape(2,4).float()
print('**Gaculate on special dim..')
print(t.max(1, keepdim=True))
print(t.min(0, keepdim=True))
print(t.sum(0, keepdim=True))
'''
**Gaculate on special dim..
torch.return_types.max(
values=tensor([[3.],
[7.]]),
indices=tensor([[3],
[3]]))
torch.return_types.min(
values=tensor([[0., 1., 2., 3.]]),
indices=tensor([[0, 0, 0, 0]]))
tensor([[ 4., 6., 8., 10.]])
'''
5.3 top-k k-th
import torch
t = torch.arange(8).reshape(2,4).float() #K大值
t2 = t.topk(3, dim=1)
print(t2)
t3 = t.topk(3, dim=1, largest=False) #K小值
print(t3)
print('*'*20)
t4 = t.kthvalue(2, dim=1)
print(t4)
'''
torch.return_types.topk(
values=tensor([[3., 2., 1.],
[7., 6., 5.]]),
indices=tensor([[3, 2, 1],
[3, 2, 1]]))
torch.return_types.topk(
values=tensor([[0., 1., 2.],
[4., 5., 6.]]),
indices=tensor([[0, 1, 2],
[0, 1, 2]]))
********************
torch.return_types.kthvalue(
values=tensor([1., 5.]),
indices=tensor([1, 1]))
'''
6. Tensor的高阶操作 6.1 where
import torch
cond = torch.tensor([[0.6,0.7],[0.8,0.4]])
a = torch.ones(2,2)
b = torch.zeros(2,2)
c = torch.where(cond>0.6, a,b)
print(c)
'''
tensor([[0., 1.],
[1., 0.]])
'''
6.2 gather
import torch table = torch.arange(4,8) index = torch.tensor([0,2,1,0,3,0]) t = torch.gather(table,dim=0,index=index) print(t) ''' tensor([4, 6, 5, 4, 7, 4]) '''
