Handling tensors

import torch

Tensor properties

Create tensor from a list or an array

example_tensor = torch.Tensor(
    [
     [[1, 2], [3, 4]], 
     [[5, 6], [7, 8]], 
     [[9, 0], [1, 2]]
    ]
)

Initialize tensors

# with ones with same shape as example tensor
torch.ones_like(example_tensor)

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

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

        [[1., 1.],
         [1., 1.]]])

# with zeros with same shape as example tensor
torch.zeros_like(example_tensor)

tensor([[[0., 0.],
         [0., 0.]],

        [[0., 0.],
         [0., 0.]],

        [[0., 0.],
         [0., 0.]]])

# with random with same shape as example tensor
torch.randn_like(example_tensor)

tensor([[[ 0.4581, -0.6200],
         [ 0.3037, -0.0642]],

        [[-1.1073, -0.1540],
         [-0.0790,  1.2929]],

        [[-1.9917,  1.1720],
         [-0.5316, -0.4773]]])

# generate tensor with shape and device
torch.randn(2, 2, device='cpu') # Alternatively, for a GPU tensor, you'd use device='cuda'

tensor([[ 0.3427, -1.3345],
        [-0.1407,  0.1086]])

Shape of tensor

example_tensor.shape

torch.Size([3, 2, 2])

print("shape[0] =", example_tensor.shape[0])
print("Rank =", len(example_tensor.shape))
print("Number of elements =", example_tensor.numel())

shape[0] = 3
Rank = 3
Number of elements = 12

Indexing tensors

example_tensor[1]

tensor([[5., 6.],
        [7., 8.]])

Get scalar value of a tensor

example_scalar = example_tensor[1, 1, 0]
example_scalar.item()

7.0

Device of the tensor

device is either cuda or cpu

example_tensor.device

device(type='cpu')

# Move tensor to a new device
#new_tensor = example_tensor.to('cuda')

Basic functions

print("Mean:", example_tensor.mean())
print("Stdev:", example_tensor.std())

Mean: tensor(4.)
Stdev: tensor(2.9848)

# average  2×2  matrix of the  3×2×2  example_tensor
example_tensor.mean(0)
# torch.mean(example_tensor, 0)

tensor([[5.0000, 2.6667],
        [3.6667, 4.6667]])