How to assign values to nn.Linear object

1

This works but the weights and bias values are random. How can I assign values to them without training over samples?

# nn.Linear
x = Variable(torch.Tensor([[1], [4]]), requires_grad=True)
# creates a wx + b layer 
f = nn.Linear(1,1)
print(f.weight, f.bias)

y = f(x)
print(y)

I have tried the following to assign values to ‘weight’ and ‘bias’

f.weight = 2.0
f.bias = 1.0


f.weight = torch.Tensor([2])
f.bias = torch.Tensor([1])


f.weight = nn.Parameter(torch.Tensor([2]))
f.bias = nn.Parameter(torch.Tensor([1]))

None seems to work.

2

If you want to operate directly on the tensors containing the parameters, you can access them using f.weight.data and f.bias.data.

I am not sure what is your context, but maybe state_dict() and load_state_dict() are the recommended methods to save and load values for some model’s parameters.

4 Likes
3

Does this still apply now that .data is deprecated alongside Variable after PyTorch 0.4?

4

As you’ve mentioned, the .data attribute and Variables are deprecated, so you could wrap the code in a with torch.no_grad() block and manipulate the parameters via e.g. param.copy_(tensor).

1 Like