Label Tensor update

[dario-coscia]

Updating LabelTensor class. Now passing to cuda/cpu and changing dtype is supported in LabelTensor. Additionally, we added the possibility to slice the tensor as:

labels = ['a', 'b']
data = torch.rand((10, 2))
tensor = LabelTensor(data, labels)
tensor_view = tensor['a']            # same as tensor.extract('a')

[dario-coscia]

@ndem0 I noticed from dir(torch.Tensor) that many implemented methods for torch lose the label. For example, let x a label tensor, then:

y1 = x.clamp(0,1)
y2 = torch.clamp(x, 0, 1)

both y1 and y2 will lose the label. Is this something that we can fix? For the first case (y1 = x.clamp(0,1)) we could think to override all the methods found in dir(torch.Tensor), for the second case I have no idea...

[dario-coscia]

@ndem0 can we merge it?

[ndem0]

@ndem0 can we merge it?

The test using [] operator is still not present. What should be the result for

label_tensor['a', 'b']

[dario-coscia]

@ndem0 can we merge it?

The test using [] operator is still not present. What should be the result for

label_tensor['a', 'b']

Added the test!