pytorch 中常用张量操作api介绍
Tensor的创建¶
从list和元组及numpy中创建¶
torch.tensor 可以从元组，list，以及 numpy 数组中创建张量
In [3]:
importtorch
importnumpyasnp
fromIPython.core.interactiveshellimport InteractiveShell
InteractiveShell.ast_node_interactivity = "all"
a =[1.1,2,3]
b = [1,2,3]
print(type(a))
c = torch.tensor(a)
d = torch.tensor(b)
c,c.dtype,c.shape
d,d.dtype,d.shape
shape = (5.00005,2)
type(shape)
torch.tensor(shape)
### 从numpy中创建
np_array = np.random.normal((2,3))
np_array
t = torch.tensor(np_array)
t.shape
t
<class 'list'>
Out[3]:
(tensor([1.1000, 2.0000, 3.0000]), torch.float32, torch.Size([3]))
Out[3]:
(tensor([1, 2, 3]), torch.int64, torch.Size([3]))
Out[3]:
tuple
Out[3]:
tensor([5.0001, 2.0000])
Out[3]:
array([4.00105954, 3.62163162])
Out[3]:
torch.Size([2])
Out[3]:
tensor([4.0011, 3.6216], dtype=torch.float64)
创建指定形状的张量¶
ones_like,zeros_like,rand_like 可以从现有的张量中去创建与之形状相同的张量
In [283]:
# 指定形状创建张量,形状可以用 元组 或许 list去指定
torch.rand([2,3])
torch.rand((2,3))
torch.randn((2,3)) # 数据服从正态分布
torch.ones([2,3])
torch.zeros(2,3) # 传入 2,3 和 (2,3) 或许[2,3] 都是一样的
torch.empty([2,3],dtype=torch.int32).random_(0,10) # 创建 形状[2,3]的张量，并填充 [0,10)之间的随机数
torch.randint(0, 10, [2,2]) #创建形状[2,2]的张量，使用数字 [0,10] 之间的数字填充
Out[283]:
tensor([[0.8735, 0.9741, 0.1682],
[0.5625, 0.8731, 0.8622]])
Out[283]:
tensor([[0.8106, 0.1381, 0.1399],
[0.1976, 0.5628, 0.9983]])
Out[283]:
tensor([[-1.7315, 0.2933, 0.9307],
[ 1.2782, 0.9797, -1.0693]])
Out[283]:
tensor([[1., 1., 1.],
[1., 1., 1.]])
Out[283]:
tensor([[0., 0., 0.],
[0., 0., 0.]])
Out[283]:
tensor([[8, 6, 7],
[9, 9, 1]], dtype=torch.int32)
Out[283]:
tensor([[7, 2],
[5, 0]])
In [284]:
torch.arange(1,5,1) # 创建一维张量， start =1 ，end= 5, step=1, 前闭后开
torch.eye(4,5) # 创建二维张量， 对角线为1， 其它位置为0
torch.full([2,3],4) # 创建一个形状为[2,3]，值全部填充为4
t = torch.tensor([1,2,3])
torch.full_like(t,6) # 以t为形状，填充6
Out[284]:
tensor([1, 2, 3, 4])
Out[284]:
tensor([[1., 0., 0., 0., 0.],
[0., 1., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 1., 0.]])
Out[284]:
tensor([[4, 4, 4],
[4, 4, 4]])
Out[284]:
tensor([6, 6, 6])
In [4]:
torch.ones_like(t)
torch.ones_like(d)
torch.zeros_like(t)
torch.rand_like(d,dtype=torch.float) #rand_like 只能生成浮点型的张量
Out[4]:
tensor([1., 1.], dtype=torch.float64)
Out[4]:
tensor([1, 1, 1])
Out[4]:
tensor([0., 0.], dtype=torch.float64)
Out[4]:
tensor([0.5891, 0.7410, 0.1948])
tensor 的属性¶
In [103]:
t=torch.ones([2,3])
t.shape # 类型 torch.Size
t.shape[1]
t.dtype # 类型 torch.dtype
t.device # 类型 torch.device
t.device.type
Out[103]:
torch.Size([2, 3])
Out[103]:
3
Out[103]:
torch.float32
Out[103]:
device(type='cpu')
Out[103]:
'cpu'
Tensor 操作¶
张量有 1200 多种操作
In [333]:
t = torch.tensor(0)
torch.is_nonzero(t) # 判断张量是否非零, 张量必须为标量（Scalar）
t = torch.tensor([[2,2,0],[4,5,6]])
torch.count_nonzero(t) #统计非0元素个数
torch.any(t == 0) # 传入的参数中有任意元素为True，则返回True
torch.numel(t) # 返回张量中所有元素的个数 ， number of elements的缩写
Out[333]:
False
Out[333]:
tensor(5)
Out[333]:
tensor(True)
Out[333]:
6
索引 切片 join 旋转¶
In [163]:
a = torch.randn([2,4,5])
a
b = torch.randn([2,3,5])
b
c = torch.cat([a,b],dim=1) # 被连接的张量必须维度相同, 除了连接的维度可以不同外，其它维度必须相同
c
c.shape
Out[163]:
tensor([[[ 0.6906, 0.2675, 0.2471, 1.0537, -0.4351],
[-0.0208, 1.1438, 0.6911, 0.9678, -0.1997],
[-0.9350, -0.6659, -1.5559, 0.1381, 0.9874],
[ 1.3853, 0.3179, 1.1886, 1.4987, 0.5392]],
[[-0.1250, -0.3843, 0.9554, -0.2116, 0.3880],
[-0.0915, 1.5961, -1.6635, -0.2928, 1.1061],
[ 0.6826, 0.3181, 0.9228, -0.7005, 0.0926],
[-0.5291, 0.8106, -0.9351, -0.6701, -1.6545]]])
Out[163]:
tensor([[[-1.3547, -0.2265, -0.8044, 0.3939, -2.0248],
[ 1.5856, 0.6330, 0.0404, -0.0779, -0.5388],
[ 0.4213, -0.4322, -0.1037, 0.4425, 0.1877]],
[[-1.0303, 0.7566, -1.0224, 0.3280, 1.1948],
[ 1.2372, -1.1186, 0.2090, -0.5584, 0.4578],
[ 0.5972, -0.2805, 0.7278, 1.4204, -0.3414]]])
Out[163]:
tensor([[[ 0.6906, 0.2675, 0.2471, 1.0537, -0.4351],
[-0.0208, 1.1438, 0.6911, 0.9678, -0.1997],
[-0.9350, -0.6659, -1.5559, 0.1381, 0.9874],
[ 1.3853, 0.3179, 1.1886, 1.4987, 0.5392],
[-1.3547, -0.2265, -0.8044, 0.3939, -2.0248],
[ 1.5856, 0.6330, 0.0404, -0.0779, -0.5388],
[ 0.4213, -0.4322, -0.1037, 0.4425, 0.1877]],
[[-0.1250, -0.3843, 0.9554, -0.2116, 0.3880],
[-0.0915, 1.5961, -1.6635, -0.2928, 1.1061],
[ 0.6826, 0.3181, 0.9228, -0.7005, 0.0926],
[-0.5291, 0.8106, -0.9351, -0.6701, -1.6545],
[-1.0303, 0.7566, -1.0224, 0.3280, 1.1948],
[ 1.2372, -1.1186, 0.2090, -0.5584, 0.4578],
[ 0.5972, -0.2805, 0.7278, 1.4204, -0.3414]]])
Out[163]:
torch.Size([2, 7, 5])
In [174]:
a = torch.randn([3,3,4])
a
torch.chunk(a,2) # 默认在第0个维度上 划分 张量
torch.chunk(a,2,1) # 划分为2 个张量，在第1个维度上划分，原张量的视图
Out[174]:
tensor([[[ 0.4793, 0.8505, -1.2706, 0.6878],
[ 0.8232, -0.4249, 1.7251, -0.8768],
[ 1.2160, 1.4610, -0.6511, -1.0861]],
[[-0.0923, 0.3947, -0.2836, 0.1569],
[ 0.1133, -0.8642, 0.2384, -0.5672],
[-0.1942, -0.0093, -2.1721, -1.0552]],
[[-0.2236, -0.2623, 0.6086, -1.0173],
[ 0.8949, -0.5393, 0.1488, -0.5906],
[ 0.0518, 0.7505, -1.0105, -0.6504]]])
Out[174]:
(tensor([[[ 0.4793, 0.8505, -1.2706, 0.6878],
[ 0.8232, -0.4249, 1.7251, -0.8768],
[ 1.2160, 1.4610, -0.6511, -1.0861]],
[[-0.0923, 0.3947, -0.2836, 0.1569],
[ 0.1133, -0.8642, 0.2384, -0.5672],
[-0.1942, -0.0093, -2.1721, -1.0552]]]),
tensor([[[-0.2236, -0.2623, 0.6086, -1.0173],
[ 0.8949, -0.5393, 0.1488, -0.5906],
[ 0.0518, 0.7505, -1.0105, -0.6504]]]))
Out[174]:
(tensor([[[ 0.4793, 0.8505, -1.2706, 0.6878],
[ 0.8232, -0.4249, 1.7251, -0.8768]],
[[-0.0923, 0.3947, -0.2836, 0.1569],
[ 0.1133, -0.8642, 0.2384, -0.5672]],
[[-0.2236, -0.2623, 0.6086, -1.0173],
[ 0.8949, -0.5393, 0.1488, -0.5906]]]),
tensor([[[ 1.2160, 1.4610, -0.6511, -1.0861]],
[[-0.1942, -0.0093, -2.1721, -1.0552]],
[[ 0.0518, 0.7505, -1.0105, -0.6504]]]))
In [196]:
a = torch.randint(0,10,[2,3,4])
a.shape
b = torch.reshape(a,(-1,4))
b
b.shape
Out[196]:
torch.Size([2, 3, 4])
Out[196]:
tensor([[1, 1, 4, 2],
[1, 6, 7, 9],
[9, 8, 9, 6],
[9, 2, 8, 0],
[2, 8, 7, 2],
[9, 7, 8, 7]])
Out[196]:
torch.Size([6, 4])
In [204]:
a = torch.randint(0,10,[2,3,4])
a
torch.split(a,2,1) # 同 torch.chunk, 原张量的视图
torch.split(a,[1,2],1) # 指定划分的比列
Out[204]:
tensor([[[6, 0, 3, 5],
[6, 1, 6, 5],
[0, 9, 4, 0]],
[[6, 0, 5, 3],
[6, 6, 1, 8],
[5, 3, 8, 9]]])
Out[204]:
(tensor([[[6, 0, 3, 5],
[6, 1, 6, 5]],
[[6, 0, 5, 3],
[6, 6, 1, 8]]]),
tensor([[[0, 9, 4, 0]],
[[5, 3, 8, 9]]]))
Out[204]:
(tensor([[[6, 0, 3, 5]],
[[6, 0, 5, 3]]]),
tensor([[[6, 1, 6, 5],
[0, 9, 4, 0]],
[[6, 6, 1, 8],
[5, 3, 8, 9]]]))
In [214]:
a = torch.randint(0,10,[2,1,3])
a
torch.squeeze(a) # 压缩所有维度为1的位置
torch.squeeze(a,[1]) # 压缩指定维度的位置
Out[214]:
tensor([[[5, 7, 2]],
[[3, 4, 4]]])
Out[214]:
tensor([[5, 7, 2],
[3, 4, 4]])
Out[214]:
tensor([[5, 7, 2],
[3, 4, 4]])
In [240]:
a = torch.randint(0,10,[2,3])
a
b = torch.randint(0,10,[2,3])
b
c = torch.randint(0,10,[2,3])
c
d = torch.stack([a,b,c],1) # 将张量给堆叠起来， 张量形状必须相同 与cat相比，该操作会扩充维度
d
d.shape
Out[240]:
tensor([[0, 0, 4],
[1, 6, 1]])
Out[240]:
tensor([[4, 4, 5],
[2, 1, 8]])
Out[240]:
tensor([[2, 2, 6],
[1, 6, 1]])
Out[240]:
tensor([[[0, 0, 4],
[4, 4, 5],
[2, 2, 6]],
[[1, 6, 1],
[2, 1, 8],
[1, 6, 1]]])
Out[240]:
torch.Size([2, 3, 3])
In [257]:
a = torch.randint(0,10,[2,3])
a
b = torch.tile(a,(2,2)) # 对a的第0维重复两次，第1维重复两次。如果传入 (2,) ,会被处理成(1,2)
b
b.shape
a.tile((2,2)) #同 torch.tile(a,(2,2))
a
Out[257]:
tensor([[3, 9, 4],
[5, 6, 6]])
Out[257]:
tensor([[3, 9, 4, 3, 9, 4],
[5, 6, 6, 5, 6, 6],
[3, 9, 4, 3, 9, 4],
[5, 6, 6, 5, 6, 6]])
Out[257]:
torch.Size([4, 6])
Out[257]:
tensor([[3, 9, 4, 3, 9, 4],
[5, 6, 6, 5, 6, 6],
[3, 9, 4, 3, 9, 4],
[5, 6, 6, 5, 6, 6]])
Out[257]:
tensor([[3, 9, 4],
[5, 6, 6]])
In [268]:
a = torch.empty([2,3,4],dtype=torch.int32).random_(0,10)
b = torch.unbind(a,1) # 拆分指定维度，并压缩， 默认拆分0维
b
b[0].shape
Out[268]:
(tensor([[6, 2, 3, 0],
[5, 7, 3, 2]], dtype=torch.int32),
tensor([[5, 3, 5, 3],
[0, 8, 6, 5]], dtype=torch.int32),
tensor([[3, 4, 4, 5],
[2, 2, 9, 4]], dtype=torch.int32))
Out[268]:
torch.Size([2, 4])
In [276]:
a = torch.empty([2,3],dtype=torch.int32).random_(0,10)
a
b = torch.ones([2,3],dtype=torch.int32)
b
torch.where(a>b,a,b) # 依次对比， 如果a>b ,给 a，否则给b
torch.where(a>2,a,2)
Out[276]:
tensor([[6, 2, 1],
[8, 0, 5]], dtype=torch.int32)
Out[276]:
tensor([[1, 1, 1],
[1, 1, 1]], dtype=torch.int32)
Out[276]:
tensor([[6, 2, 1],
[8, 1, 5]], dtype=torch.int32)
Out[276]:
tensor([[6, 2, 2],
[8, 2, 5]], dtype=torch.int32)
In [278]:
torch.manual_seed(1) # 随机种子设定
Out[278]:
<torch._C.Generator at 0x240d615ab70>
In [299]:
torch.randperm(4) # [0,4) 创建一维张量，元素个数维4个，数字从[0,4)随机打乱
# 数据打乱
data = torch.tensor([10, 20, 30, 40, 50])
perm = torch.randperm(data.size(0))
shuffled_data = data[perm]
shuffled_data
Out[299]:
tensor([3, 0, 1, 2])
Out[299]:
tensor([30, 50, 10, 40, 20])
In [304]:
t = torch.randint(0,10,[2,3]);
t.shape
t.size()
t.size(1)
Out[304]:
torch.Size([2, 3])
Out[304]:
torch.Size([2, 3])
Out[304]:
3
In [322]:
a = torch.randn(2, 3)
a
a.max()
a.argmax() # 返回张量中最大值索引
a.argmax(1) # 指定维度查找，返回所在维度的索引
Out[322]:
tensor([[-0.3013, -0.7432, -0.6355],
[ 0.4731, -0.1942, 0.3147]])
Out[322]:
tensor(0.4731)
Out[322]:
tensor(3)
Out[322]:
tensor([0, 0])
In [329]:
a = torch.randn(1,1)
a
a.item() # 获取张量中的值， 当张量中仅有一个值的时候有效
Out[329]:
tensor([[-0.3149]])
Out[329]:
-0.3148927390575409
张量的索引选择， 易混乱部分¶
In [354]:
t = torch.rand([2,3])
t
t[0][-1]
t[0,-1]
Out[354]:
tensor([[0.1808, 0.2796, 0.3273],
[0.3835, 0.2156, 0.6563]])
Out[354]:
tensor(0.3273)
Out[354]:
tensor(0.3273)
In [344]:
# 列表索引
x = torch.tensor([[[10, 20, 30], [40, 50, 60]]])
x[[0, 0]] #作用在第0维度，列表[0,0]，这意味着在维0上选择第0个元素两次
x[[0, 0], [0, 1]] #分别作用在维0和维1上，选择两个元素， 分别是0维上的 0 和 1维度上的 0 与0维度上的0和1维度上的1
x[0, [0, 1], [0, 1]] #作用在三个维度，分别是0维上的 0 和 1维度上的 0 和2维度上的0 与 0维度上的0和1维度上的1和2维度上的1
Out[344]:
tensor([[[10, 20, 30],
[40, 50, 60]],
[[10, 20, 30],
[40, 50, 60]]])
Out[344]:
tensor([[10, 20, 30],
[40, 50, 60]])
Out[344]:
tensor([10, 50])
In [348]:
x4 = torch.tensor([[10, 20, 30], [40, 50, 60]])
x4[[False, True], [False, True, True]]
Out[348]:
tensor([50, 60])
In [351]:
x = torch.tensor([[10, 20, 30], [40, 50, 60]])
x[torch.tensor([[0, 1], [1, 0]])] #选择x的第一个维度，并重新排序
Out[351]:
tensor([[[10, 20, 30],
[40, 50, 60]],
[[40, 50, 60],
[10, 20, 30]]])
In [352]:
importtorch
x = torch.tensor([[10, 20, 30], [40, 50, 60]]) # 形状 (2, 3)
index = torch.tensor([[False, True, True], [False, True, True]]) # 形状 (2, 3)
x[index]
Out[352]:
tensor([20, 30, 50, 60])
In [353]:
x = torch.tensor([[10, 20, 30], [40, 50, 60]])
index = torch.tensor([False, True])
x[index]
Out[353]:
tensor([[40, 50, 60]])
知识是我们已知的
也是我们未知的
基于已有的知识之上
我们去发现未知的
由此,知识得到扩充
我们获得的知识越多
未知的知识就会更多
因而,知识扩充永无止境
浙公网安备 33010602011771号