Pytorch入门：Tensor创建

Tensor即张量，是深度学习中常用的数据形式。

Pytorch官方文档：torch — PyTorch 1.9.1 documentation

Pytorch中文文档：主页 - PyTorch中文文档

Tensor创建函数

常用tensor创建函数如下：

""" 根据传入数据创建张量 """
torch.tensor(data, dtype=None, device=None, requires_grad=False)

""" 根据传入数据创建张量，torch.Tensor()等同于torch.FloatTensor() """ 
torch.Tensor(data, dtype=None, device=None, requires_grad=False)



""" 根据size创建值在[0, 1)范围的随机张量 """ 
torch.rand(*size, dtype=None, device=None, requires_grad=False)

""" 根据size创建值为1的张量 """ 
torch.ones(*size, dtype=None, device=None, requires_grad=False)

""" 根据size创建值为0的张量 """ 
torch.zeros(*size, dtype=None, device=None, requires_grad=False)

""" 根据size创建值为value的张量 """ 
torch.full([*size], value, dtype=None, device=None, requires_grad=False)



""" 根据传入tensor的size创建值为1的张量 """ 
torch.ones_like(tensor, dtype=None, device=None, requires_grad=False)

""" 根据传入tensor的size创建值为0的张量 """ 
torch.zeros_like(tensor, dtype=None, device=None, requires_grad=False)

""" 根据传入tensor的size创建值为value的张量 """ 
torch.full_like(tensor, value, dtype=None, device=None, requires_grad=False)

dtype参数：数据类型。可省略。

device参数：该张量存放于cpu或gpu，如device='cpu'或device='cuda:0'。可省略，默认为cpu。

requires_grad参数：该张量是否需要在计算中保留对应的梯度信息，训练过程的tensor需设置为True，测试阶段可设置为False。可省略，默认为False。

下表列出了Tensor类与其对应dtype详情：

Data type	dtype	Tensor types
32-bit floating point	torch.float32 或 torch.float	torch.*.FloatTensor
64-bit floating point	torch.float64 或 torch.double	torch.*.DoubleTensor
16-bit floating point	torch.float16 或 torch.half	torch.*.HalfTensor
8-bit integer (unsigned)	torch.uint8	torch.*.ByteTensor
8-bit integer (signed)	torch.int8	torch.*.CharTensor
16-bit integer (signed)	torch.int16 或 torch.short	torch.*.ShortTensor
32-bit integer (signed)	torch.int32 或 torch.int	torch.*.IntTensor
64-bit integer (signed)	torch.int64 或 torch.long	torch.*.LongTensor

Tensor信息查看

print('tensor: ', tensor)
print('type: ', tensor.type())
print('size: ', tensor.size())
print('dim: ', tensor.dim())
print('device: ', tensor.device)
print('dtype: ', tensor.dtype)
print('requires_grad: ', tensor.requires_grad)

Tensor创建示例

0维Tensor（标量）

t = torch.tensor(1)

"""
tensor:  tensor(1)
type:  torch.LongTensor
size:  torch.Size([])
dim:  0
device:  cpu
dtype:  torch.int64
requires_grad:  False
"""

1维Tensor（向量）

t = torch.tensor([1, 2, 3])

"""
tensor:  tensor([1, 2, 3])
type:  torch.LongTensor
size:  torch.Size([3])
dim:  1
device:  cpu
dtype:  torch.int64
requires_grad:  False
"""

t = torch.Tensor([1, 2, 3])

"""
tensor:  tensor([1., 2., 3.])
type:  torch.FloatTensor
size:  torch.Size([3])
dim:  1
device:  cpu
dtype:  torch.float32
requires_grad:  False
"""

注：torch.Tensor()为torch.FloatTensor()的别名

２维Tensor（矩阵）

t = torch.tensor([[1, 2, 3], [4, 5, 6]])

"""
tensor:  tensor([[1, 2, 3],
                 [4, 5, 6]])
type:  torch.LongTensor
size:  torch.Size([2, 3])
dim:  2
device:  cpu
dtype:  torch.int64
requires_grad:  False
"""

t = torch.rand(2, 3)

"""
tensor:  tensor([[0.3600, 0.5245, 0.3553],
                 [0.4554, 0.8006, 0.8296]])
type:  torch.FloatTensor
size:  torch.Size([2, 3])
dim:  2
device:  cpu
dtype:  torch.float32
requires_grad:  False
"""

t2 = torch.full_like(t, 2)

"""
tensor:  tensor([[2., 2., 2.],
                 [2., 2., 2.]])
type:  torch.FloatTensor
size:  torch.Size([2, 3])
dim:  2
device:  cpu
dtype:  torch.float32
requires_grad:  False
"""

n维Tensor（n>=3）（张量）

t = torch.full([2, 3, 5], 2, dtype=torch.float, device='cuda:0', requires_grad=False)

"""
tensor:  tensor([[[2., 2., 2., 2., 2.],
                  [2., 2., 2., 2., 2.],
                  [2., 2., 2., 2., 2.]],
                 [[2., 2., 2., 2., 2.],
                  [2., 2., 2., 2., 2.],
                  [2., 2., 2., 2., 2.]]], device='cuda:0')
type:  torch.cuda.FloatTensor
size:  torch.Size([2, 3, 5])
dim:  3
device:  cuda:0
dtype:  torch.float32
requires_grad:  False
"""

Tensor的GPU、CPU切换

可通过以下语句判断当前环境下GPU是否可用：

if torch.cuda.is_available():
    print('GPU is available.')

若GPU可用，则可将张量转移到GPU：

t = torch.rand(2, 3)

t = t.cuda()    # 将张量转移到GPU
print(t.device) # cuda:0
 
t = t.cpu()     # 将张量转移到CPU
print(t.device) # cpu

也可以通过to方法进行转移：

device = 'cuda' if torch.cuda.is_available() else 'cpu'

t = torch.rand(2, 3)
t = t.to(device)
print(t.device)    # cuda:0