How to move a Torch Tensor from CPU to GPU and vice versa?

A torch tensor defined on CPU can be moved to GPU and vice versa. For high-dimensional tensor computation, the GPU utilizes the power of parallel computing to reduce the compute time.

High-dimensional tensors such as images are highly computation-intensive and takes too much time if run over the CPU. So, we need to move such tensors to GPU.

Syntax

• To move a torch tensor from CPU to GPU, following syntax/es are used −

Tensor.to("cuda:0")
or
Tensor.to(cuda)

And,

Tensor.cuda()

• To move a torch tensor from GPU to CPU, the following syntax/es are used −

Tensor.to("cpu")

And,

Tensor.cpu()

Let's take a couple of examples to demonstrate how a tensor can be moved from CPU to GPU and vice versa.

Note − I have provided two different outputs for each program. One output for the systems having CPU only and the other output for system having GPU along with CPU.

Example 1

# Python program to move a tensor from CPU to GPU
# import torch library
import torch

# create a tensor
x = torch.tensor([1.0,2.0,3.0,4.0])
print("Tensor:", x)

# check tensor device (cpu/cuda)
print("Tensor device:", x.device)

# Move tensor from CPU to GPU
# check CUDA GPU is available or not
print("CUDA GPU:", torch.cuda.is_available())
if torch.cuda.is_available():
   x = x.to("cuda:0")
   # or x=x.to("cuda")
print(x)

# now check the tensor device
print("Tensor device:", x.device)

Output 1 − When GPU is not available

Tensor: tensor([1., 2., 3., 4.])
Tensor device: cpu
CUDA GPU: False
tensor([1., 2., 3., 4.])
Tensor device: cpu

Output 2 − When GPU is available

Tensor: tensor([1., 2., 3., 4.])
Tensor device: cpu
CUDA GPU: True
tensor([1., 2., 3., 4.], device='cuda:0')
Tensor device: cuda:0

Example 2

# Python program to move a tensor from CPU to GPU
# import torch library
import torch

# create a tensor on CPU
x = torch.tensor([1.0,2.0,3.0,4.0])
print("Tensor:", x)
print("Tensor device:", x.device)

# Move tensor from CPU to GPU
if torch.cuda.is_available():
   x = x.cuda()
print(x)
# now check the tensor device
print("Tensor device:", x.device)

Output 1 − If GPU is not available

Tensor: tensor([1., 2., 3., 4.])
Tensor device: cpu
tensor([1., 2., 3., 4.])
Tensor device: cpu

Output 2 − If GPU is available

Tensor: tensor([1., 2., 3., 4.])
Tensor device: cpu
tensor([1., 2., 3., 4.], device='cuda:0')
Tensor device: cuda:0

Example 3

# Python program to move a tensor from GPU to CPU
# import torch library
import torch

# create a tensor on GPU
if torch.cuda.is_available():
   x = torch.tensor([1.0,2.0,3.0,4.0], device = "cuda")
print("Tensor:", x)
print("Tensor device:", x.device)

# Move tensor from GPU to CPU
x = x.to("cpu")
# x = x.cpu()
print(x)

# Now check the tensor device
print("Tensor device:", x.device)

Output 1 − If GPU is not available

Tensor: tensor([1., 2., 3., 4.])
Tensor device: cpu
tensor([1., 2., 3., 4.])
Tensor device: cpu

Output 2 − If GPU is available

Tensor: tensor([1., 2., 3., 4.], device='cuda:0')
Tensor device: cuda:0
tensor([1., 2., 3., 4.])
Tensor device: cpu