创建Tensor
从numpy导入
In [62]: a=np.array([2,3.3])
In [63]: torch.from_numpy(a)
0ut[63]: tensor([2.0000, 3.3000], dtype=torch.float64)
//从NUMPY导入的FLOAT其实是DOUBLE类型
In [65]: a=np.ones([2,3])
In [66]: torch.from_numpy(a)
0ut[66]:tensor([[1.,1.],[1., 1.,1.]],dtype=torch.float64)
从list导入
In [67]: torch.tensor([2.,3.2])
Out[67]: tensor([2. 0000, 3.2000])
In [68]: torch.FloatTensor([2., 3.2])
//容易引起歧义
0ut[68]: tensor([2.0000,3.2000])
In [69]: torch.tensor([[2., 3.2], [1.,22.396]])
0ut[69]: tensor([[2.0000,3.2000],[1.0000, 22.3000]])
未初始化
Torch.empty()
Torch.FloatTensor(d1,d2,d3)
Torch.IntTensor(d1,d2,d3)
//随机初始化的张量元素会非常大或者非常小
//因此可能会导致NAN或者Infinity的问题
设置默认数据类型
In [74]: torch.tensor([1.2,3]).type()
0ut[74]: 'torch. FloatTensor'
In [75]: torch.set_default_tensor_type(torch. DoubleTensor)
In [76]: torch.tensor([1.2,3]).type()
0ut[76]: 'torch. DoubleTensor'
随机初始化
a = torch.rand(3,3)//在[0,1)中随机生成3*3的向量
//[0,10)就*10即可
torch.rand_like(a)//随机生成与a的shape相同的向量
torch.randint(d1,d2)//随机生成d1*d2的整数向量
torch.normal(mean=torch.full([10],0), std=torch.arange(1, 0, -0.1))
//生成以0为均值,方差从1到0逐渐减小[1,0.9,0.8...,0.1],维度为1,长度为10的向量
有目标地初始化
full
torch.full([10],0)//生成长度为10,但是都为0的向量
torch.full([],1)//生成值为1的标量
arange
In [96]: torch.arange(0,10)
0ut[96]: tensor([0, 1, 2, 3, 4, 5,6, 7, 8, 9])
In [97]: torch.arange(0,10,2)
0ut[97]: tensor([0, 2, 4, 6, 8])
linspace/logspace
In [99]: torch.linspace(0,10, steps=4 )
0ut[99]: tensor([0. 0000,3.3333,6.6667,10. 0000])
//从0到10切割成4个数字
In [103]: torch.logspace(0,-1, steps=10)
0ut[103]: tensor([1.0000,0.7743, 0.5995, 0.4642, 0.3594, 0.2783, 0.2154, 0.1668,0.1292, 0.1000])
//10^0到10^-1之间切割成10个数字
ones/zeros/eye
In [105]: torch.ones(3,3)
0ut[105] :
tensor([[1.,1.,1.],
[1.,1.,1.],
[1.,1.,1.]])
//全1向量
In [107]: torch.eye(3,4)
0ut[107] :
tensor([[1.,0.,0.,0.],
[0.,1.,0.,0.],
[0.,0.,1.,0.]])
//对角线为1的向量
In [110]: torch.zeros(3,3)
0ut[110]:
tensor([[0.,0.,0.],
[0.,0.,0.],
[0.,0.,0.]])
//全0向量
torch.ones_like()
随机打散
randperm
In [127]: torch.randperm(10)
0ut[127]: tensor([1, 5, 4, 2, 0, 6, 3, 9, 7, 8])