fix bug when using cuda in ubuntu

[WorldHellooo]

The Class Variable's init() doesn't return a cuda Variable because the 'data' would be released later. This bug results in free(): invalid pointer error.

[apaszke]

How is that different from the previous solution (except that it always produces volatile variables which is incorrect)?

[WorldHellooo]

The previous solution returns a cpu Variable even if the USE_CUDA is true. As a result, the data is in the cpu while the net model is in the gpu, causing a invalid pointer error.

[apaszke]

No, it doesn't. The data is converted to CUDA

[WorldHellooo]

Could you tell me what's wrong with my experiment?

[apaszke]

Ah I now see what's wrong. Variables ignore __init__. In fact now I'd consider modifying these classes bad practice and the example should be modified. However your PR doesn't fix the problem.

[WorldHellooo]

Ok, it doesn't matters.

[chsasank]

Hi,

@WorldHellooo does this PR fix #66 ?
@apaszke What changes do you recommend? I'll make those changes.

[WorldHellooo]

In my test, it did fix #66.

[apaszke]

It might fix it, but it's not a correct patch. We should use this hack but call type

[chsasank]

Cool, I'll create a PR like this

use_cuda = torch.cuda.is_available()
dtype = torch.cuda.FloatTensor if use_cuda else torch.FloatTensor
.
.
.

var = Variable(data).type(dtype)