Tensors

Tensors

Tensors behave almost exactly the same way in PyTorch as they do in Torch.

Create a tensor of size (5 x 7) with uninitialized memory:

import torch
a = torch.empty(5, 7, dtype=torch.float)

Initialize a double tensor randomized with a normal distribution with mean=0, var=1:

a = torch.randn(5, 7, dtype=torch.double)
print(a)
print(a.size())

Out:

tensor([[ 0.6654, -0.2213, -0.2664,  0.0911,  0.6865,  0.2613, -1.0404],
        [-1.0137, -1.4412, -1.0891,  0.4997,  1.1697, -0.8738, -1.6117],
        [ 0.6082, -1.3122, -0.6268, -0.6429,  1.5538, -1.5196, -0.1233],
        [ 0.5462, -0.1450,  0.0067,  1.1419,  1.2654, -1.2371,  0.5570],
        [-0.0183, -0.4094,  0.5403,  1.3002, -0.1093, -0.0864,  0.0845]],
       dtype=torch.float64)
torch.Size([5, 7])

Note

torch.Size is in fact a tuple, so it supports the same operations

Inplace / Out-of-place

The first difference is that ALL operations on the tensor that operate in-place on it will have an _postfix. For example, add is the out-of-place version, and add_ is the in-place version.

a.fill_(3.5)
# a has now been filled with the value 3.5

b = a.add(4.0)
# a is still filled with 3.5
# new tensor b is returned with values 3.5 + 4.0 = 7.5

print(a, b)

Out:

tensor([[3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000],
        [3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000],
        [3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000],
        [3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000],
        [3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000, 3.5000]],
       dtype=torch.float64) tensor([[7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000],
        [7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000],
        [7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000],
        [7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000],
        [7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000, 7.5000]],
       dtype=torch.float64)

Some operations like narrow do not have in-place versions, and hence, .narrow_ does not exist. Similarly, some operations like fill_ do not have an out-of-place version, so .fill does not exist.

Zero Indexing

Another difference is that Tensors are zero-indexed. (In lua, tensors are one-indexed)

b = a[0, 3]  # select 1st row, 4th column from a

Tensors can be also indexed with Python’s slicing

扫描二维码关注公众号，回复： 2940027 查看本文章

b = a[:, 3:5]  # selects all rows, 4th column and  5th column from a

No camel casing

The next small difference is that all functions are now NOT camelCase anymore. For example indexAdd is now called index_add_

x = torch.ones(5, 5)
print(x)

Out:

tensor([[1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1.]])

z = torch.empty(5, 2)
z[:, 0] = 10
z[:, 1] = 100
print(z)

Out:

tensor([[ 10., 100.],
        [ 10., 100.],
        [ 10., 100.],
        [ 10., 100.],
        [ 10., 100.]])

x.index_add_(1, torch.tensor([4, 0], dtype=torch.long), z)
print(x)

Out:

tensor([[101.,   1.,   1.,   1.,  11.],
        [101.,   1.,   1.,   1.,  11.],
        [101.,   1.,   1.,   1.,  11.],
        [101.,   1.,   1.,   1.,  11.],
        [101.,   1.,   1.,   1.,  11.]])

Numpy Bridge

Converting a torch Tensor to a numpy array and vice versa is a breeze. The torch Tensor and numpy array will share their underlying memory locations, and changing one will change the other.

Converting torch Tensor to numpy Array

a = torch.ones(5)
print(a)

Out:

tensor([1., 1., 1., 1., 1.])

b = a.numpy()
print(b)

Out:

[1. 1. 1. 1. 1.]

a.add_(1)
print(a)
print(b)    # see how the numpy array changed in value

Out:

tensor([2., 2., 2., 2., 2.])
[2. 2. 2. 2. 2.]

Converting numpy Array to torch Tensor

import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a)
print(a)
print(b)  # see how changing the np array changed the torch Tensor automatically

Out:

[2. 2. 2. 2. 2.]
tensor([2., 2., 2., 2., 2.], dtype=torch.float64)

All the Tensors on the CPU except a CharTensor support converting to NumPy and back.

CUDA Tensors

CUDA Tensors are nice and easy in pytorch, and transfering a CUDA tensor from the CPU to GPU will retain its underlying type.

# let us run this cell only if CUDA is available
if torch.cuda.is_available():

    # creates a LongTensor and transfers it
    # to GPU as torch.cuda.LongTensor
    a = torch.full((10,), 3, device=torch.device("cuda"))
    print(type(a))
    b = a.to(torch.device("cpu"))
    # transfers it to CPU, back to
    # being a torch.LongTensor

Out:

<class 'torch.Tensor'>