python|【pytorch】pytorch基础深度学习|numpy|大数据|java

一、张量数据类型 1.1 pytorch与python数据类型对比

python	pytorch
Int	IntTensor of size()
float	FloatTensor of size()
Int array	IntTensor of size [d1,d2,…]
Float array	FloatTensor of size [d1,d2,…]
string	ont-hot or Embedding(Word2Vec,glove)

由于PyTorch不是一个完备的语言库，它是面向数据计算的GPU加速库，因此没有内建对string的支持，在PyTorch中可用以下两种方式来表示string：

One-hot
Embedding
- Word2vec
- glove

1.2 pytorch内建的数据类型

Data tyoe	dtype	CPU tensor	GPU tensor
32-bit floating point	torch.float32 or torch.float	`torch.FloatTensor`	`torch.cuda.FloatTensor`
64-bit floating point	torch.float64 or torch.double	`torch.DoubleTensor`	`torch.cuda.DoubleTensor`
16-bit floating point	torch.float16 or torch.half	N/A	`torch.cuda.HalfTensor`
8-bit integer (unsigned)	torch.uint8	`torch.ByteTensor`	`torch.cuda.ByteTensor`
8-bit integer (signed)	torch.int8	`torch.CharTensor`	`torch.cuda.CharTensor`
16-bit integer (signed)	torch.int16 or torch.short	`torch.ShortTensor`	`torch.cuda.ShortTensor`
32-bit integer (signed)	torch.int32 or torch.int	`torch.IntTensor`	`torch.cuda.IntTensor`
64-bit integer (signed)	torch.int64 or torch.long	`torch.LongTensor`	`torch.cuda.LongTensor`

Tpye check

Tensor.type()
type(Tensor)
isinstance(Tensor,torch.FloatTensor)

>>> a = torch.randn(2,3) >>> a.type() 'torch.FloatTensor' >>> type(a) >>> isinstance(a,torch.FloatTensor) True >>> isinstance(a,torch.cuda.DoubleTensor) False >>> data = https://www.it610.com/article/a.cuda() #x.cuda()返回一个gpu的引用>>> isinstance(data,torch.cuda.DoubleTensor) >>> True

标量：Dimension 0/rank 0，Deep Learning中常用于loss表示

>>> torch.tensor(2.2)#.tensor接受的是数据的内容 tensor(2.2000) >>> a = torch.tensor(2.2) >>> a.shape torch.Size([]) >>> len(a.shape) 0 >>> a.dim() 0 >>> a.size() torch.Size([])

Dimension 1/rank 1，Deep Learning中常用于Bias表示

>>> torch.tensor([1.1]) tensor([1.1000]) >>> torch.tensor([1.1,2.2]) tensor([1.1000, 2.2000]) >>> torch.FloatTensor(2) #接受的是size，random初始化 tensor([2.2758e-07, 4.5682e-41]) >>> data = https://www.it610.com/article/np.ones(2)>>> data array([1., 1.]) >>> torch.from_numpy(data) tensor([1., 1.], dtype=torch.float64)

Dimension 2

>>> a = torch.randn(2,3) #FloatTensor(2,3) >>> a tensor([[-0.1680,0.4534, -0.4045], [-1.0437, -0.4634,0.7432]]) >>> a.shape torch.Size([2, 3]) >>> a.size(0) 2 >>> a.size(1) 3 >>> a.shape[0] 2 >>> a.shape[1] 3

【python|【pytorch】pytorch基础】Dim 3:RNN Input with Batch

>>> a=torch.rand(1,2,3) >>> a tensor([[[0.2226, 0.0342, 0.1301], [0.6371, 0.6930, 0.9356]]]) >>> a.shape torch.Size([1, 2, 3]) >>> a[0] #[2,3] tensor([[0.2226, 0.0342, 0.1301], [0.6371, 0.6930, 0.9356]]) >>> list(a.shape) [1, 2, 3]

Dim4:CNN[b,c,h,w]
numbel是指tensor占用内存的数量

>>> a = torch.rand(2,3,28,28) >>> a tensor([[[[0.7190, 0.8762, 0.3667,..., 0.8682, 0.5834, 0.7012], [0.4110, 0.5633, 0.1516,..., 0.6877, 0.1930, 0.9480], [0.0063, 0.8593, 0.4722,..., 0.4012, 0.8891, 0.0254], ..., ..., [0.3267, 0.8081, 0.5329,..., 0.3658, 0.9325, 0.6759], [0.4113, 0.8107, 0.9934,..., 0.2609, 0.1763, 0.5233], [0.7673, 0.3748, 0.0287,..., 0.0348, 0.0529, 0.8054]]]]) >>> >>> a.shape torch.Size([2, 3, 28, 28]) >>> a.numel() 4704 >>> a.dim() #len(a.shape) 4 >>> a=torch.tensor(1) >>> a.dim() 0

创建Tensor
Import form array

>>> a=np.array([2,3,3]) >>> torch.from_numpy(a) tensor([2, 3, 3]) >>> a=np.ones([2,3,3]) >>> torch.from_numpy(a) tensor([[[1., 1., 1.], [1., 1., 1.], [1., 1., 1.]],[[1., 1., 1.], [1., 1., 1.], [1., 1., 1.]]], dtype=torch.float64)

Import from List

torch.tensor([2.,3.2]) tensor([2.0000, 3.2000]) torch.FloatTensor([2.,3.2]) #这种方式不推荐使用，接受具体的数据推荐使用torch,tensor tensor([2.0000, 3.2000]) torch.Tensor([2.,3.2]) tensor([2.0000, 3.2000]) torch.tensor([[2.,3.2],[1.,22.3]]) tensor([[ 2.0000,3.2000], [ 1.0000, 22.3000]])

**torch.FloatTensor和torch.Tensor基本是一样的，都是接受书的维度shape创建Tensor，而小写的torch.tensor是接受具体的数据**

uninitialized

Torch.empty(2,3)
Torch.FloatTensor(d1,d2,d3)
Torch.IntTensor(d1,d2,d3)

注意未初始化的api在使用的时候，若后续没有赋值操作覆盖，将会使得随机初始化的值变的非常大或者非常小(troch.nan/inf)
随机初始化

torch.rand
torch.rand_like
torch.randint
torch.randn:正态分布
torch.normal:自定义分布

>>> torch.rand(3,3) tensor([[0.3628, 0.4771, 0.5067], [0.6593, 0.6323, 0.9157], [0.5882, 0.6289, 0.4311]]) >>> a=torch.rand(3,3) >>> torch.rand_like(a) tensor([[0.5168, 0.9998, 0.1509], [0.6104, 0.5265, 0.7917], [0.3084, 0.9142, 0.0541]]) >>> torch.randint(1,10,[3,3]) tensor([[2, 2, 6], [3, 9, 1], [4, 5, 4]]) >>> torch.normal(mean=torch.full([10],0),std=torch.arange(1,0,-0.1)) tensor([-2.1002, -0.2133,0.9746,0.6781,0.3725,0.6669,0.4720,0.7872, 0.0643,0.0143])

set default type

>>> torch.tensor([1.2,3]).type() 'torch.FloatTensor' #pytorch默认的数据类型是FloatTensor >>> torch.set_default_tensor_type(torch.DoubleTensor) >>> torch.tensor([1.2,3]).type() 'torch.DoubleTensor'#增强学习一般使用double

二、torch常用函数在使用Tensor时，我们首先要掌握如何使用Tensor来定义不同数据类型的变量。和Numpy差不多，PyTorch中的Tensor也有自己的数据类型定义方式，常用的如下：
1.张量Tensors

torch.is_tensor(obj) #判断是否为张量，如果是pytorch张量，则返回True
torch.is_storage(obj) #判断是否为pytorch Storage，如何是，则返回True
torch.set_default_tensor_type(t)
torch.numel(input)->int
torch.set_printoptions(precision=None, threshold=None, edgeitems=None, linewidth=None, profile=None) #设置打印选项。完全参考自 Numpy。

2.创建操作 Creation Ops torch.eye

torch.eye(n, m=None, out=None)->Tensor

返回一个2维张量(单位矩阵)，对角线数字为1，其它位置为0
参数说明：

n (int) – 行数
m (int, 可选) – 列数.如果为None,则默认为n
out (Tensor,可选) - 输出张量

>>> torch.eye(4) tensor([[1., 0., 0., 0.], [0., 1., 0., 0.], [0., 0., 1., 0.], [0., 0., 0., 1.]])

torch.zero

torch.zeros(*sizes, out=None) → Tensor

返回一个全0的张量，形状由可变参数sizes定义。

>>> torch.zeros(2,3) tensor([[0., 0., 0.], [0., 0., 0.]]) >>> torch.zeros(5) tensor([0., 0., 0., 0., 0.])

torch.ones

torch.ones(*sizes, out=None) → Tensor #返回一个全为1的张量，形状由可变参数sizes定义。

>>> torch.ones(2,3) tensor([[1., 1., 1.], [1., 1., 1.]]) >>> torch.ones(5) tensor([1., 1., 1., 1., 1.])

torch.from_numpy

torch.from_numpy(ndarray) → Tensor

将numpy.ndarray转换为Tensor。返回的张量tensor和numpy的ndarray共享同一内存空间。修改一个会导致另外一个也被修改。返回的张量不能调整大小。

>>> a= np.array([1,2,4]) >>> t = torch.from_numpy(a) >>> t tensor([1, 2, 4]) >>> t[1]=10 >>> a array([ 1, 10,4])

torch.linspace

torch.linspace(start, end, steps=100, out=None) → Tensor #返回start和end之间长度为steps的一维张量

参数说明：

start (float) – 点集的起始值
end (float) – 点集的最终值
steps (int) – 在start 和 end间的采样数,即返回多少个数
out (Tensor, 可选的) – 结果张量

>>> torch.linspace(1.0,10,steps=5) tensor([ 1.0000,3.2500,5.5000,7.7500, 10.0000]) >>> torch.linspace(-10,10,steps=5,out=None) tensor([-10.,-5.,0.,5.,10.])

torch.logspace

torch.logspace(start, end, steps=100, out=None) → Tensor

返回一个1维张量，包含在区间\(10^{start}\) 和\(10^{end}\)上以对数刻度均匀间隔的steps个点。输出1维张量的长度为steps,参数说明同torch.linspace

>>> torch.logspace(start=-10,end=10,steps=5) tensor([1.0000e-10, 1.0000e-05, 1.0000e+00, 1.0000e+05, 1.0000e+10]) >>> torch.logspace(start=0.1,end=1.0,steps=5) tensor([ 1.2589,2.1135,3.5481,5.9566, 10.0000])

torch.FloatTensor

此变量用于生成数据类型为浮点型的Tensor，传递给torch.FloatTensor的参数可以是一个列表，也可以是一个维度值。torch.IntTensor与此类似。

>>> import torch >>> a = torch.FloatTensor(2,3) >>> print(a) -0.11710.0000 -0.1171 0.00000.00000.0000 [torch.FloatTensor of size 2x3] >>> b = torch.FloatTensor([2,3,4,5]) >>> print(b) 2 3 4 5 [torch.FloatTensor of size 4]

torch.rand

用于生成数据类型为浮点型且维度指定的随机Tensor，和在Numpy中使用numpy.rand生成随机数的方法类似，随机生成的浮点数据在0~1区间均匀分布。

torch.rand(*sizes, out=None) → Tensor
返回一个张量，填充在[0,1]区间的一组均匀分布随机数。 Tensor的形状由变量sizes定义。

>>> torch.rand(4) tensor([0.0662, 0.7079, 0.4197, 0.2257]) >>> torch.rand(2,3,out=None) tensor([[0.8174, 0.8959, 0.2713], [0.5343, 0.0275, 0.7712]]) >>> torch.rand(2,3)

torch.randn

torch.randn(*sizes, out=None) → Tensor

用于生成数据类型为浮点型且维度指定的随机Tensor，和在Numpy中使用numpy.randn生成随机数的方法类似，随机生成的浮点数的取值满足均值为0，方差为1的正态分布。

>>> torch.randn(4) tensor([-1.4524,0.9949, -1.4038,0.8059]) >>> torch.randn(2,3) tensor([[ 0.9186,1.0356,1.1103], [-2.0057, -0.9032,0.6453]])

torch.randperm

torch.randperm(n, out=None) → LongTensor

输入参数n(int) – 上限(独占),即最大值，返回一个从0 到n -1的随机整数排列，随机打散。
random.shuffle

>>> torch.randperm(4) tensor([3, 1, 2, 0])>>> a=torch.rand(2,3) >>> b=torch.rand(2,2) >>> idx=torch.randperm(2) >>> idx tensor([1, 0]) >>> idx tensor([1, 0]) >>> a,b (tensor([[0.8094, 0.8389, 0.5666], [0.6812, 0.5959, 0.4951]]), tensor([[0.4863, 0.5345], [0.1246, 0.2468]])) >>> a[idx] tensor([[0.6812, 0.5959, 0.4951], [0.8094, 0.8389, 0.5666]]) >>> b[idx] tensor([[0.1246, 0.2468], [0.4863, 0.5345]])

torch.arange

torch.arange(start, end, step=1, out=None) → Tensor([start; end))

返回一个1维张量，长度为floor((end?start)/step),floor代表向下取整。包含从start到end，以step为步长的一组序列值(默认步长为1)。

>>> torch.arange(1,4) tensor([1, 2, 3]) >>> torch.arange(1,2.5,0.5) tensor([1.0000, 1.5000, 2.0000])

torch.range

torch.range(start, end, step=1, out=None) → Tensor([start; end]),现在torch.range将要废弃，

返回一个1维张量，长度为floor((end?start)/step)+1，其中floor代表向下取整数。从start开始，end为结尾，以step为步长的一组值。 step 是两个值之间的间隔，即 Xi+1=Xi+step

>>> import torch >>> a = torch.range(2,8,1) __main__:1: UserWarning: torch.range is deprecated in favor of torch.arange and will be removed in 0.5. Note that arange generates values in [start; end), not [start; end]. >>> print(a) tensor([2., 3., 4., 5., 6., 7., 8.])

torch.full

>>> torch.full([2,3],7) tensor([[7., 7., 7.], [7., 7., 7.]]) >>> torch.full([],7) #给标量赋值 tensor(7.) >>> torch.full([1],7) tensor([7.]) >>> torch.full([2],7) tensor([7., 7.])

3.Indexing, Slicing, Joining, Mutating Ops

dim 0 first
select first/last N
select by steps
select by specific index
...

dim 0 first

>>> a = torch.rand(4,3,28,28) >>> a[0].shape torch.Size([3, 28, 28]) >>> a[0,0].shape torch.Size([28, 28]) >>> a[0,0,2,4] tensor(0.2409)

select first/last N

>>> a[:2].shape torch.Size([2, 3, 28, 28]) >>> a[:2,:1,:,:].shape torch.Size([2, 1, 28, 28]) >>> a[:2,1:,:,:].shape torch.Size([2, 2, 28, 28]) >>> a[:2,-1:,:,:].shape torch.Size([2, 1, 28, 28])

select by steps

>>> a[:,:,0:28:2,:].shape torch.Size([4, 3, 14, 28]) >>> a[:,:,0:28:2,0:14].shape torch.Size([4, 3, 14, 14]) >>> a[:,:,0:28:2,2:14].shape torch.Size([4, 3, 14, 12])

通用形式：start : end : step
select by specific index
torch.index_select

torch.index_select(input, dim, index, out=None) → Tensor
参数：

input (Tensor) – 输入张量

dim (int) – 索引的轴

index (LongTensor) – 包含索引下标的一维张量

out (Tensor, optional) – 目标张量

沿着指定维度对输入进行切片，取index中指定的相应项(index为一个LongTensor)，然后返回到一个新的张量，返回的张量与原始张量_Tensor_有相同的维度(在指定轴上)。
注意：返回的张量不与原始张量共享内存空间。

>>> x=torch.randn(3,4) >>> x tensor([[ 0.4878, -0.7477, -0.2496,1.1835], [-1.8953,0.3530,0.1122,0.9137], [ 0.4686,0.5230, -1.1191, -1.0911]]) >>> indices = torch.LongTensor([0,2]) >>> indices tensor([0, 2]) >>> torch.index_select(x,1,indices) tensor([[ 0.4878, -0.2496], [-1.8953,0.1122], [ 0.4686, -1.1191]]) >>> torch.index_select(x,1,indices) tensor([[ 0.4878, -0.2496], [-1.8953,0.1122], [ 0.4686, -1.1191]]) >>> torch.index_select(x,0,indices) tensor([[ 0.4878, -0.7477, -0.2496,1.1835], [ 0.4686,0.5230, -1.1191, -1.0911]])

...

>>> a.shape torch.Size([4, 3, 28, 28]) >>> a[...].shape torch.Size([4, 3, 28, 28]) >>> a[0,...].shape#等同于a[0] torch.Size([3, 28, 28]) >>> a[:,1,...].shape torch.Size([4, 28, 28]) >>> a[0,...,::2].shape#等同于a[0,:,:,::2] torch.Size([3, 28, 14]) >>> a[...,:2].shape torch.Size([4, 3, 28, 2])

select by mask
torch.masked_select

torch.masked_select(input, mask, out=None) → Tensor

input (Tensor) – 输入张量

mask (ByteTensor) – 掩码张量，包含了二元索引值

out (Tensor, optional) – 目标张量

根据掩码张量mask中的二元值，取输入张量中的指定项( mask为一个 ByteTensor)，将取值返回到一个新的1D张量，
张量 mask须跟input张量有相同数量的元素数目，但形状或维度不需要相同。注意：返回的张量不与原始张量共享内存空间。

>>> x = torch.randn(3,4) >>> x tensor([[ 0.3132, -0.4580, -0.3642,0.7394], [-0.2821,1.9086, -0.9687,1.6009], [ 0.9800, -0.8546, -0.8855, -0.3807]]) >>> mask = x.ge(0.5) >>> mask tensor([[False, False, False,True], [False,True, False,True], [ True, False, False, False]]) >>> torch.masked_select(x,mask) tensor([0.7394, 1.9086, 1.6009, 0.9800]) >>> torch.masked_select(x,mask).shape torch.Size([4])

select by flatten index

>>> src = https://www.it610.com/article/torch.tensor([[4,3,5],[6,7,8]])>>> torch.take(src,torch.tensor([0,2,5])) tensor([4, 5, 8])

torch.nonzero

torch.nonzero(input, out=None) → LongTensor
参数:

input (Tensor) – 源张量

out (LongTensor, optional) – 包含索引值的结果张量

返回一个包含输入input中非零元素索引的张量。输出张量中的每行包含输入中非零元素的索引。如果输入input有n维，则输出的索引张量output的形状为 z x n, 这里 z 是输入张量input中所有非零元素的个数。

>>> torch.nonzero(torch.Tensor([1,1,1,0,1])) #返回一维张量中非零的索引值 tensor([[0], [1], [2], [4]]) #输入张量input有2(n)维，非零元素有4(z)个，所以输出张量output的shape为4× 2 >>> torch.nonzero(torch.Tensor([[0.6,0.0,0.0,0.0],[0.0,0.4,0.0,0.0],[0.0,0.0,1.2,0.0],[0.0,0.0,5,0.0]])) tensor([[0, 0], [1, 1], [2, 2], [3, 2]]) #表明input的第4行的index2的为非零数据

维度变换

view/reshape(view保证numel不变即可)
Squeeze/unsqueeze(删减/增加)
Transpose/t/permute
Expand/repeat

torch.view/torch.reshape
view操作需保证数据有实际意义，不然会将数据造成破坏，数据的存储、维度顺序非常重要

>>> a=torch.rand(4,1,28,28)#4张灰度图片 >>> a.view(4,28*28)#(4,1*28*28)将channel和像素值合并在一起=》[4,784]适合于全连接层 tensor([[0.2301, 0.9408, 0.3547,..., 0.9387, 0.0988, 0.9476], [0.2724, 0.6440, 0.0037,..., 0.7575, 0.1136, 0.7190], [0.6347, 0.9259, 0.4316,..., 0.8456, 0.2670, 0.6662], [0.0801, 0.3157, 0.4126,..., 0.4852, 0.2193, 0.8381]]) >>> a.view(4,28*28).shape torch.Size([4, 784]) >>> a.view(4*28,28).shape torch.Size([112, 28]) >>> a.view(4*1,28,28).shape torch.Size([4, 28, 28]) #(4*1,28*28),指关注featuremap >>> b=a.view(4,784) >>> b.view(4,28,28,1) #logic Bug,将维度信息丢失掉了 tensor([[[[0.2301], [0.9408], [0.3547], ..., [0.4852], [0.2193], [0.8381]]]])

torch.unsqueeze

torch.unsqueeze(input, dim, out=None)
参数:

tensor (Tensor) – 输入张量

dim (int) – 插入维度的索引

out (Tensor, optional) – 结果张量

返回一个新的张量，对输入的指定位置插入维度。注意：返回张量与输入张量共享内存，所以改变其中一个的内容会改变另一个。如果dim为负，则将会被转化dim+input.dim()+1，指定位置在 1[-a.dim()-1,a.dim()+1]=>[-5,5]之间变化。

>>> a.shape torch.Size([4, 1, 28, 28]) >>> a.unsqueeze(0).shape torch.Size([1, 4, 1, 28, 28]) >>> a.unsqueeze(-1).shape torch.Size([4, 1, 28, 28, 1]) >>> a.unsqueeze(-4).shape torch.Size([4, 1, 1, 28, 28]) >>> >>> a.unsqueeze(-5).shape torch.Size([1, 4, 1, 28, 28]) >>> a.unsqueeze(5).shape Traceback (most recent call last): File "", line 1, in IndexError: Dimension out of range (expected to be in range of [-5, 4], but got 5)>>> a=torch.tensor([1.2,2.3]) >>> a.unsqueeze(-1)# [2,1] tensor([[1.2000], [2.3000]]) >>> a.unsqueeze(0) #[1,2] tensor([[1.2000, 2.3000]])

example:

>>> b=torch.rand(32) #bias相当于给每个channel上的所有像素增加一个偏置 >>> f=torch.rand(4,32,14,14) >>> b=b.unsqueeze(1).unsqueeze(2).unsqueeze(0) >>> b.shape torch.Size([1, 32, 1, 1])

torch.squeeze

>>> b.shape torch.Size([1, 32, 1, 1]) >>> b.squeeze().shape torch.Size([32]) >>> b.squeeze(0).shape torch.Size([32, 1, 1]) >>> b.shape torch.Size([1, 32, 1, 1]) >>> b.squeeze(-1).shape torch.Size([1, 32, 1]) >>> b.squeeze(1).shape#32是不能被挤压的，只能挤压1的维度 torch.Size([1, 32, 1, 1]) >>> b.squeeze(-4).shape torch.Size([32, 1, 1])

torch.unbind

torch.unbind(tensor, dim=0)[source]
参数:

tensor (Tensor) – 输入张量

dim (int) – 删除的维度

移除指定维后，返回一个元组，包含了沿着指定维切片后的各个切片
torch.expand/torch.repeat

Expand:broadcasting【不会主动复制数据，只在需要的时候，这是推荐的一种方式; 进行操作的时候需要前后dim一致，只能从dim1->dimN; -1表示dim不变】
Repeat:memory copied(给出的参数p为在相应的维度上拷贝p次)

###expand >>> a=torch.rand(4,32,14,14) >>> b.shape torch.Size([1, 32, 1, 1]) >>> b.expand(4,32,14,14).shape torch.Size([4, 32, 14, 14]) >>> b.expand(-1,32,-1,-1).shape torch.Size([1, 32, 1, 1]) >>> b.expand(-1,32,-1,-4).shape#-4是bug,在最新版的facebook已经被修复了 torch.Size([1, 32, 1, -4]) ##repeat >>> b.shape torch.Size([1, 32, 1, 1]) >>> b.repeat(4,32,1,1).shape torch.Size([4, 1024, 1, 1]) >>> b.repeat(4,1,1,1).shape torch.Size([4, 32, 1, 1]) >>> b.repeat(4,1,32,32).shape torch.Size([4, 32, 32, 32])

torch.t

torch.t(input, out=None) → Tensor
参数:

input (Tensor) – 输入张量

out (Tensor, optional) – 结果张量

输入一个矩阵(2维张量)，并转置0, 1维。可以被视为函数transpose(input, 0, 1)的简写函数,但是torch.t只适用于2D。
torch.transpose

torch.transpose(input, dim0, dim1, out=None) → Tensor

input (Tensor) – 输入张量

dim0 (int) – 转置的第一维

dim1 (int) – 转置的第二维

返回输入矩阵input的转置。交换维度dim0和dim1。输出张量与输入张量共享内存，所以改变其中一个会导致另外一个也被修改。高维矩阵的转置常用torch.transpose

>>> a=torch.rand(4,3,32,32) >>> a1.shape torch.Size([4, 32, 32, 3]) >>> a1 = a.transpose(1,3).view(4,3*32*32).view(4,3,32,32)#[b,c,h,w]->[b,w,h,c]->[b,c,w,h]维度信息与原来的存储信息不一致，报错 Traceback (most recent call last): File "", line 1, in RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead. >>> a1 = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,3,32,32) #contiguous将数据变为连续，同上，维度污染，错误 >>> a2 = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,32,32,3).transpose(1,3) >>> a1.shape,a2.shape (torch.Size([4, 3, 32, 32]), torch.Size([4, 3, 32, 32])) >>> torch.all(torch.eq(a,a1)) tensor(False) >>> torch.all(torch.eq(a,a2)) tensor(True)

permute

permute()参数信息是原来Tensor的维度信息

[b,h,w,c]是numpy存储图片的格式，只有这个格式才能导出numpy

>>> a=torch.rand(4,3,28,28) #[b,c,h,w] >>> a.transpose(1,3).shape torch.Size([4, 28, 28, 3]) >>> b=torch.rand(4,3,28,32) >>> b.transpose(1,3).shape #[b,c,h,w]->[b,w,h,c] torch.Size([4, 32, 28, 3]) >>> b.transpose(1,3).shape #[b,w,h,c]->[b,c,h,w] torch.Size([4, 32, 28, 3]) >>> b.permute(0,2,3,1).shape #[b,c,h,w]->[b,h,w,c] torch.Size([4, 28, 32, 3])