Why would we use to() method in pytorch?

Why would we use to(device) method in pytorch?

torch.Tensor.to is multipurpose method.

Not only you can do type conversion, but it can also do CPU to GPU tensor move and GPU to CPU tensor move:

tensor = torch.randn(2, 2)  
print(tensor)
tensor = tensor.to(torch.float64) 
print(tensor) #dtype=torch.float64
tensor = tensor.to("cuda") 
print(tensor) #device='cuda:0', dtype=torch.float64)
tensor = tensor.to("cpu") 
print(tensor) #dtype=torch.float64
tensor = tensor.to(torch.float32) 
print(tensor) # won't print dtype=torch.float32 since it is by default

Since CPU and GPU are different kind memories, there must be a way they communicate. This is why we have to("cuda"), and to("cpu") that we call on tensor.

Usually when you load training datasets (images):

• you download them from URL (like MNIST http://deeplearning.net/data/mnist/mnist.pkl.gz)
• unpack them
• convert them to numpy arrays
• convert numpy arrays to tensors (as this is fast)
• move them to GPU for training .to("cuda")

You can create tensors and move them to GPU like this.

torch.zeros(1000).to("cuda")

But there is a trick, sometimes you can even load them directly to GPU without messing the CPU.

torch.zeros(1000, device="gpu")

.to(device)

.to() can be used to copy a tensor to any available device (CPU, GPU):

x = torch.randn(3)

x_g0 = x.to('cuda:0')
x_c = x.to('cpu')

It has aliases .cuda() and .cpu().

.to(dtype)

When given a dtype as an arument .to() acts as a casting method:

x_d = x.to(torch.double)

It has dtype-named aliases .double(), .float(), .int() etc