Hi,
I love the PyTorch library so far, but I wanted to note one little annoyance that I encountered -- something that could be improved in future. For instance, when I convert numpy arrays to torch tensors, it disregards the original NumPy type and creates torch.DoubleTensor /torch.cuda.DoubleTensors. I think it would make more sense to inherit the array type float, int, etc. to create the tensors.

Here's a simple example:

import numpy as np
import torch

y = np.array([[1., 2., 3.],
              [4., 5., 6.]])

torch.from_numpy(y)

Output:

1  2  3
4  5  6
[torch.DoubleTensor of size 2x3]