CODEWITHSUNDEEP
pythonnumpydeep-learningpytorchnumpy-ndarray
Lanka
Lanka

Reputation: 357

how to create empty parameter in in pytorch tensor variable

from torch import FloatTensor

def new_parameter(*size): #1024
    out = torch.nn.Parameter(FloatTensor(*size), requires_grad=True)
    torch.nn.init.xavier_normal_(out)
    return out

at = new_parameter(1024, 1)

output is

Parameter containing:
tensor([[ 0.0203],
        [-0.0043],
        [-0.0386],
        ...,
        [-0.0084],
        [-0.0289],
        [-0.0188]], requires_grad=True)

similarway we can create

bt=torch.randn((1024,1),requires_grad=True)

output also same

tensor([[-1.5478],
        [ 1.5060],
        [ 0.1580],
        ...,
        [ 0.9754],
        [ 0.1699],
        [ 0.2062]], requires_grad=True)

are there any differences in a tensor variable the above two ways? please explain the above code in simply

Upvotes: 1

Views: 1130

Answers (1)

Ivan
Ivan

Reputation: 40738

The first method will initialize a random float tensor, then wrap it with nn.Parameter. Which is generally used to register than tensor as a parameter to a nn.Module (not seen here). A utility function nn.init.xavier_normal_ is then applied on that parameter to initialize its values.

The second method only initializes a random float tensor.

Upvotes: 1

Related Questions