• 【pytorch】pytorch基础


    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)
      <class 'torch.Tensor'>
      >>> isinstance(a,torch.FloatTensor)
      True
      >>> isinstance(a,torch.cuda.DoubleTensor)
      False
      >>> data = 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 = 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
      

      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

    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startend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})(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)
     
    -0.1171  0.0000 -0.1171
     0.0000  0.0000  0.0000
    [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0n -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startend,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 = 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inputn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 "<stdin>", line 1, in <module>
    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dim0dim1。 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 "<stdin>", line 1, in <module>
    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])
  • 相关阅读:
    leetcode 买卖股票的最佳时机3
    leetcode 买卖股票的最佳时机Ⅱ
    leetcode 最长有效括号
    C++中的%lld和%I64d区别
    Ural 1095 Nikifor 3 思维+同余性质的利用
    博弈基础
    ural 1091. Tmutarakan Exams
    容斥原理
    一些易错的地方
    codeforces911D Inversion Counting 求逆序数+小trick
  • 原文地址:https://www.cnblogs.com/lyszyl/p/12161819.html
Copyright © 2020-2023  润新知